@wendellresearch-ai/benchwarmers 0.2.1 → 0.3.0

package/metrics/quota.py

@@ -0,0 +1,383 @@
+"""Quota records: what a provider says is left on a connection, and when it resets.
+
+    python metrics/quota.py              # every record still in force
+    python metrics/quota.py --selftest   # each reset format and provider profile; no calls
+
+RELEASE-PLAN section 2, "Quotas". The built-in client (metrics/client.py)
+hands every response's headers here, and every 429's body; what they state
+becomes one record per connection, model, metric and window:
+
+    {"connection": "groq", "alias": "...", "metric": "requests", "window": "day",
+     "limit": 1000, "remaining": 998, "reset_at": "2026-10-24T00:00:12+00:00",
+     "source": "headers (groq)", "observed_at": "2026-10-23T19:01:30+00:00"}
+
+kept in runs/quotas.json (not config: it expires, and is nobody's decision),
+by connection and then by model, since Groq, Cerebras and OpenAI meter each
+model apart. Pacing (metrics/pacing.py) reads them back: calls sent since a
+record was observed are taken off its `remaining`, and when nothing is left
+the next call waits for `reset_at` or is refused. A record stops counting at
+its reset, so a connection is never held for want of traffic to observe one.
+
+ONE HEADER NAME, DIFFERENT WINDOWS
+----------------------------------
+`x-ratelimit-limit-requests` is per minute at OpenAI and per day at Groq;
+Cerebras names its windows (`-requests-day`, `-tokens-minute`); OpenRouter
+sends bare `X-RateLimit-*` only on a 429, inside the body's metadata as well,
+and its message says which window ("free-models-per-day"); Anthropic's
+`anthropic-ratelimit-*` are per minute; Gemini sends no headers but its 429
+body names the exhausted quota and the retry delay. PROFILES holds what each
+provider means, picked by the endpoint's host; any other host is read the
+OpenAI way with the window left "unknown", which still enforces, since the
+reset is absolute.
+
+RESETS COME IN FIVE SPELLINGS
+-----------------------------
+Durations (`6m0s`, `2m59.56s`, `20ms`, `17s`), float seconds (Cerebras'
+`33011.38`), epoch seconds or milliseconds (OpenRouter's `1761264000000`),
+RFC 3339 (Anthropic) and HTTP dates (Retry-After). Each becomes an absolute
+time the moment the response arrives: "in 6m0s" means nothing to whoever
+reads the record later.
+
+metrics/learn.py still learns from the error TEXT, for either backend: a
+stated limit, a context window, a cooldown. A 429 body that states a limit
+with a known window (Gemini's "50 a day") is passed on to it here, so the
+limit is learned (and enforced) like one read from a message.
+"""
+
+from __future__ import annotations
+
+import datetime as dt
+import email.utils
+import json
+import pathlib
+import re
+import sys
+import time
+import urllib.parse
+
+import locked
+
+ROOT = pathlib.Path(__file__).resolve().parent.parent
+from paths import DATA  # noqa: E402  -- the install's data folder (metrics/paths.py)
+FILE = DATA / "runs" / "quotas.json"
+KEEP_S = 7 * 86400                     # a record with no reset is shown this long, never enforced
+
+# host (or a parent domain of it) -> (profile, window of x-ratelimit-*-requests / -tokens
+# when the header does not name one)
+PROFILES = {
+    "api.openai.com": ("openai", {"requests": "minute", "tokens": "minute"}),
+    "api.groq.com": ("groq", {"requests": "day", "tokens": "minute"}),
+    "api.cerebras.ai": ("cerebras", {}),
+    "openrouter.ai": ("openrouter", {}),
+    "api.anthropic.com": ("anthropic", {}),
+    "api.sambanova.ai": ("sambanova", {"requests": "minute", "tokens": "minute"}),
+    "generativelanguage.googleapis.com": ("gemini", {}),
+}
+WINDOWS = {"minute": 60, "hour": 3600, "day": 86400}
+OPENAI_STYLE = re.compile(r"^x-ratelimit-(limit|remaining|reset)-(requests|tokens)(?:-(minute|hour|day))?$")
+BARE = re.compile(r"^x-ratelimit-(limit|remaining|reset)$")
+ANTHROPIC = re.compile(r"^anthropic-ratelimit-(requests|tokens|input-tokens|output-tokens)-(limit|remaining|reset)$")
+DURATION = re.compile(r"^(?:(\d+(?:\.\d+)?)h)?(?:(\d+(?:\.\d+)?)m(?!s))?(?:(\d+(?:\.\d+)?)s)?(?:(\d+(?:\.\d+)?)ms)?$")
+
+
+def profile_for(base: str) -> tuple[str, dict]:
+    """The profile for an endpoint's base URL; ("generic", {}) for any other host."""
+    host = (urllib.parse.urlsplit(base or "").hostname or "").lower()
+    for known, prof in PROFILES.items():
+        if host == known or host.endswith("." + known):
+            return prof
+    return "generic", {}
+
+
+def reset_time(value, now: float) -> float | None:
+    """An absolute epoch time from any of the five spellings, or None."""
+    raw = str(value or "").strip()
+    if not raw:
+        return None
+    if (m := DURATION.match(raw)) and any(m.groups()):
+        h, mi, s, ms = (float(g) if g else 0.0 for g in m.groups())
+        return now + h * 3600 + mi * 60 + s + ms / 1000
+    if re.match(r"^\d{4}-\d{2}-\d{2}T", raw):
+        try:
+            return dt.datetime.fromisoformat(raw.replace("Z", "+00:00")).timestamp()
+        except ValueError:
+            return None
+    try:
+        n = float(raw)
+    except ValueError:
+        try:                                           # an HTTP date
+            return email.utils.parsedate_to_datetime(raw).timestamp()
+        except (TypeError, ValueError):
+            return None
+    if n >= 1e12:                                      # epoch milliseconds
+        return n / 1000
+    if n >= 1e9:                                       # epoch seconds
+        return n
+    return now + n                                     # seconds from now
+
+
+def _iso(t: float | None) -> str | None:
+    return None if t is None else dt.datetime.fromtimestamp(t, dt.timezone.utc).isoformat(timespec="seconds")
+
+
+def _int(v) -> int | None:
+    try:
+        return int(float(str(v).strip()))
+    except (TypeError, ValueError):
+        return None
+
+
+def _window_hint(text: str) -> str:
+    """OpenRouter says which window ran out in its message, not its headers."""
+    low = (text or "").lower()
+    for word, window in (("per-day", "day"), ("per day", "day"), ("per-min", "minute"), ("per min", "minute")):
+        if word in low:
+            return window
+    return "unknown"
+
+
+def from_headers(headers, profile: tuple[str, dict], now: float, status: int = 200,
+                 hint: str = "", source: str = "headers") -> list[dict]:
+    """The records a response's headers state. `hint`: the error message, which
+    names OpenRouter's window."""
+    name, defaults = profile
+    h = {str(k).lower(): v for k, v in (headers.items() if hasattr(headers, "items") else [])}
+    fields: dict[tuple[str, str], dict] = {}
+    for key, value in h.items():
+        if m := OPENAI_STYLE.match(key):
+            what, metric, window = m.group(1), m.group(2), m.group(3) or defaults.get(m.group(2), "unknown")
+        elif m := BARE.match(key):
+            what, metric, window = m.group(1), "requests", _window_hint(hint)
+        elif m := ANTHROPIC.match(key):
+            metric, what, window = m.group(1), m.group(2), "minute"
+        else:
+            continue
+        fields.setdefault((metric, window), {})[what] = value
+    out = []
+    for (metric, window), f in fields.items():
+        rec = {"metric": metric, "window": window, "limit": _int(f.get("limit")),
+               "remaining": _int(f.get("remaining")), "reset_at": _iso(reset_time(f.get("reset"), now)),
+               "source": f"{source} ({name})", "observed_at": _iso(now)}
+        if rec["limit"] is not None or rec["remaining"] is not None:
+            out.append(rec)
+    if status == 429:
+        wait = None
+        for key, scale in (("retry-after-ms", 0.001), ("retry-after", 1.0)):
+            if h.get(key):
+                raw = str(h[key]).strip()
+                try:
+                    wait = float(raw) * scale
+                except ValueError:
+                    t = reset_time(raw, now)
+                    wait = None if t is None else t - now
+                break
+        if wait and wait > 0:
+            out.append({"metric": "requests", "window": "retry-after", "limit": None, "remaining": 0,
+                        "reset_at": _iso(now + wait), "source": f"retry-after ({name})",
+                        "observed_at": _iso(now)})
+    return out
+
+
+def _error_objects(body):
+    """Each `error` dict in a 429 body: an object, or Gemini's one-item list."""
+    for item in (body if isinstance(body, list) else [body]):
+        if isinstance(item, dict) and isinstance(item.get("error"), dict):
+            yield item["error"]
+
+
+def from_body(body, profile: tuple[str, dict], now: float) -> list[dict]:
+    """The records a 429 body states: Gemini's QuotaFailure and RetryInfo, and
+    OpenRouter's copy of its rate-limit headers in `metadata.headers`."""
+    out = []
+    for err in _error_objects(body):
+        meta = err.get("metadata") if isinstance(err.get("metadata"), dict) else {}
+        if isinstance(meta.get("headers"), dict):
+            out += from_headers(meta["headers"], profile, now, hint=str(err.get("message") or ""),
+                                source="429 body")
+        details = [d for d in err.get("details") or [] if isinstance(d, dict)]
+        delay = next((reset_time(d.get("retryDelay"), now) for d in details if d.get("retryDelay")), None)
+        for d in details:
+            for v in d.get("violations") or []:
+                if not isinstance(v, dict) or not v.get("quotaId"):
+                    continue
+                qid = str(v["quotaId"])
+                window = next((w for w in ("day", "hour", "minute") if f"per{w}" in qid.lower()), "unknown")
+                metric = "tokens" if "token" in (qid + str(v.get("quotaMetric") or "")).lower() else "requests"
+                # No retry delay on a per-minute quota: the minute is the wait.
+                reset = delay if delay else (now + WINDOWS[window] if window == "minute" else None)
+                out.append({"metric": metric, "window": window, "limit": _int(v.get("quotaValue")),
+                            "remaining": 0, "reset_at": _iso(reset), "quota": qid,
+                            "source": f"429 body ({profile[0]})", "observed_at": _iso(now)})
+    return out
+
+
+def _ts(iso: str | None) -> float | None:
+    try:
+        return dt.datetime.fromisoformat(iso).timestamp() if iso else None
+    except (TypeError, ValueError):
+        return None
+
+
+def store(connection: str, alias: str, records: list[dict], now: float | None = None) -> None:
+    """Keep the newest record per metric and window, and drop what has expired."""
+    now = time.time() if now is None else now
+
+    def change(data: dict) -> None:
+        mine = data.setdefault(connection, {}).setdefault(alias, {})
+        for r in records:
+            mine[f"{r['metric']}/{r['window']}"] = {"connection": connection, "alias": alias, **r}
+        for conn in list(data):
+            if conn.startswith("_") or not isinstance(data[conn], dict):
+                continue
+            for a in list(data[conn]):
+                recs = data[conn][a]
+                for k in list(recs):
+                    reset, seen = _ts(recs[k].get("reset_at")), _ts(recs[k].get("observed_at")) or 0
+                    if (reset is not None and reset < now - 86400) or (reset is None and seen < now - KEEP_S):
+                        del recs[k]
+                if not recs:
+                    del data[conn][a]
+            if not data[conn]:
+                del data[conn]
+        data["_why"] = ("What providers' rate-limit headers and 429 bodies said was left, per connection"
+                        " and model (metrics/quota.py). Expires; nobody's decision.")
+    locked.update(FILE, change, ensure_ascii=False)
+
+
+def observe(alias: str, connection: str, base: str, status: int, headers, body=None) -> list[dict]:
+    """Record what one response states; returns the records. Called by the
+    built-in client on every reply and every HTTP error."""
+    now, profile = time.time(), profile_for(base)
+    hint = " ".join(str(e.get("message") or "") for e in _error_objects(body)) if status == 429 else ""
+    records = from_headers(headers, profile, now, status, hint)
+    if status == 429 and body is not None:
+        records += from_body(body, profile, now)
+    if records:
+        store(connection or "?", alias, records, now)
+        if status == 429:
+            _learn_limits(alias, records)
+    return records
+
+
+def _learn_limits(alias: str, records: list[dict]) -> None:
+    """A 429 that states a per-minute or per-day limit: learned, as learn.py
+    learns one stated in an error message (only ever tighter)."""
+    axes = {("requests", "minute"): "rpm", ("requests", "day"): "rpd",
+            ("tokens", "minute"): "tpm", ("tokens", "day"): "tpd"}
+    found = {axes[(r["metric"], r["window"])]: r["limit"] for r in records
+             if (r["metric"], r["window"]) in axes and isinstance(r.get("limit"), int) and r["limit"] > 0}
+    if found:
+        try:
+            import learn
+            learn.note_limits(alias, found, "quota", "; ".join(r["source"] for r in records)[:200])
+        except Exception:                              # noqa: BLE001  -- learning never breaks a call
+            pass
+
+
+def live(connection: str | None = None, alias: str | None = None, now: float | None = None) -> list[dict]:
+    """Records whose reset is still ahead, for one connection and model, or all."""
+    now = time.time() if now is None else now
+    out = []
+    for conn, by_alias in locked.read_json(FILE).items():
+        if conn.startswith("_") or not isinstance(by_alias, dict) or (connection and conn != connection):
+            continue
+        for a, recs in by_alias.items():
+            if (alias and a != alias) or not isinstance(recs, dict):
+                continue
+            out += [r for r in recs.values() if isinstance(r, dict) and (_ts(r.get("reset_at")) or 0) > now]
+    return out
+
+
+def _selftest() -> None:
+    now = 1_761_264_000.0                              # 2025-10-24T00:00:00Z
+    # Each reset spelling.
+    assert reset_time("6m0s", now) == now + 360
+    assert abs(reset_time("2m59.56s", now) - (now + 179.56)) < 1e-6     # Groq
+    assert reset_time("20ms", now) == now + 0.02                        # OpenAI
+    assert reset_time("1h2m3s", now) == now + 3723
+    assert reset_time("17s", now) == now + 17                           # Gemini's retryDelay
+    assert abs(reset_time("33011.382867", now) - (now + 33011.382867)) < 1e-6   # Cerebras
+    assert reset_time("1761350400000", now) == 1761350400.0             # OpenRouter, epoch ms
+    assert reset_time("1761350400", now) == 1761350400.0                # epoch s
+    assert reset_time("2025-10-24T00:01:00Z", now) == now + 60          # Anthropic, RFC 3339
+    assert reset_time("Fri, 24 Oct 2025 00:00:30 GMT", now) == now + 30  # an HTTP date
+    assert reset_time("", now) is None and reset_time("soon", now) is None
+
+    def one(recs, metric, window):
+        got = [r for r in recs if r["metric"] == metric and r["window"] == window]
+        assert len(got) == 1, (metric, window, recs)
+        return got[0]
+    # OpenAI: unsuffixed means per minute.
+    r = from_headers({"x-ratelimit-limit-requests": "500", "x-ratelimit-remaining-requests": "499",
+                      "x-ratelimit-reset-requests": "120ms", "x-ratelimit-limit-tokens": "30000",
+                      "x-ratelimit-remaining-tokens": "29000", "x-ratelimit-reset-tokens": "2s"},
+                     profile_for("https://api.openai.com/v1"), now)
+    assert one(r, "requests", "minute")["remaining"] == 499 and one(r, "tokens", "minute")["limit"] == 30000
+    # Groq: the same names, but requests are per day.
+    r = from_headers({"X-Ratelimit-Limit-Requests": "1000", "X-Ratelimit-Remaining-Requests": "0",
+                      "X-Ratelimit-Reset-Requests": "2m59.56s"}, profile_for("https://api.groq.com/openai/v1"), now)
+    g = one(r, "requests", "day")
+    assert g["remaining"] == 0 and g["reset_at"] == _iso(now + 179.56) and g["source"] == "headers (groq)", g
+    # Cerebras names its windows.
+    r = from_headers({"x-ratelimit-limit-requests-day": "14400", "x-ratelimit-remaining-requests-day": "14399",
+                      "x-ratelimit-reset-requests-day": "33011.38", "x-ratelimit-limit-tokens-minute": "60000",
+                      "x-ratelimit-remaining-tokens-minute": "59000", "x-ratelimit-reset-tokens-minute": "11.4"},
+                     profile_for("https://api.cerebras.ai/v1"), now)
+    assert one(r, "requests", "day")["limit"] == 14400 and one(r, "tokens", "minute")["remaining"] == 59000
+    # Anthropic: RFC 3339 resets, per minute.
+    r = from_headers({"anthropic-ratelimit-requests-limit": "50", "anthropic-ratelimit-requests-remaining": "0",
+                      "anthropic-ratelimit-requests-reset": "2025-10-24T00:00:45Z"},
+                     profile_for("https://api.anthropic.com/v1"), now)
+    assert one(r, "requests", "minute")["reset_at"] == _iso(now + 45)
+    # OpenRouter: bare headers, epoch ms, window from the message; and retry-after on a 429.
+    r = from_headers({"X-RateLimit-Limit": "50", "X-RateLimit-Remaining": "0", "X-RateLimit-Reset": "1761350400000",
+                      "Retry-After": "30"}, profile_for("https://openrouter.ai/api/v1"), now, 429,
+                     hint="Rate limit exceeded: free-models-per-day")
+    assert one(r, "requests", "day")["reset_at"] == _iso(1761350400.0)
+    assert one(r, "requests", "retry-after")["reset_at"] == _iso(now + 30)
+    # Retry-After is only a quota on a 429.
+    assert from_headers({"Retry-After": "30"}, ("generic", {}), now, 503) == []
+    # Any other host: read the OpenAI way, window unknown.
+    r = from_headers({"x-ratelimit-remaining-requests": "3", "x-ratelimit-reset-requests": "10s"},
+                     profile_for("http://127.0.0.1:9/v1"), now)
+    assert one(r, "requests", "unknown")["source"] == "headers (generic)"
+    assert from_headers({"content-type": "application/json"}, ("generic", {}), now) == []
+    # Gemini's 429 body (a one-item list through the OpenAI-compatible endpoint).
+    gem = [{"error": {"code": 429, "status": "RESOURCE_EXHAUSTED", "details": [
+        {"@type": "type.googleapis.com/google.rpc.QuotaFailure", "violations": [
+            {"quotaMetric": "generativelanguage.googleapis.com/generate_content_free_tier_requests",
+             "quotaId": "GenerateRequestsPerDayPerProjectPerModel-FreeTier", "quotaValue": "50"}]},
+        {"@type": "type.googleapis.com/google.rpc.RetryInfo", "retryDelay": "17s"}]}}]
+    g = one(from_body(gem, profile_for("https://generativelanguage.googleapis.com/v1beta/openai"), now),
+            "requests", "day")
+    assert g["limit"] == 50 and g["remaining"] == 0 and g["reset_at"] == _iso(now + 17), g
+    tok = {"error": {"details": [{"violations": [{"quotaId": "GenerateContentInputTokensPerModelPerMinute-FreeTier",
+                                                  "quotaValue": "250000"}]}]}}
+    g = one(from_body(tok, ("gemini", {}), now), "tokens", "minute")
+    assert g["reset_at"] == _iso(now + 60), g                          # no delay: the minute is the wait
+    # OpenRouter's 429 body carries its headers in metadata.
+    orb = {"error": {"message": "Rate limit exceeded: free-models-per-min.", "code": 429,
+                     "metadata": {"headers": {"X-RateLimit-Limit": "20", "X-RateLimit-Remaining": "0",
+                                              "X-RateLimit-Reset": "1761264060000"}}}}
+    g = one(from_body(orb, profile_for("https://openrouter.ai/api/v1"), now), "requests", "minute")
+    assert g["limit"] == 20 and g["source"] == "429 body (openrouter)", g
+    assert from_body("<html>busy</html>", ("generic", {}), now) == []
+    print("ok")
+
+
+def main() -> None:
+    sys.stdout.reconfigure(encoding="utf-8", errors="replace")
+    if "--selftest" in sys.argv:
+        _selftest()
+        return
+    recs = live()
+    if not recs:
+        print(f"no quota record in force ({FILE})")
+        return
+    for r in sorted(recs, key=lambda r: (r["connection"], r["alias"], r["metric"], r["window"])):
+        print(f"  {r['connection']:<16} {r['alias']:<30} {r['metric']:<8} {r['window']:<11}"
+              f" {r.get('remaining')!s:>7} of {r.get('limit')!s:<7} until {r.get('reset_at')}  ({r['source']})")
+
+
+if __name__ == "__main__":
+    main()

package/metrics/route.py

@@ -76,9 +76,11 @@ task. A poor score never does. A model that did badly once on one task shape
 has earned less confidence, not exile.
 
 Cost, separately. A model listed under `ad_hoc` in config/model-roles.json is
-expensive or tightly capped and is never seated automatically — `--with` is the
-only way onto the panel. Staleness would otherwise pick it first, since it is
-exactly the model nobody uses casually.
+expensive or tightly capped and is never seated automatically: `--with` is the
+only way onto this router's panel. Staleness would otherwise pick it first,
+since it is exactly the model nobody uses casually. A collection that lists it
+as a member (or merger) still calls it on every request; that is a choice made
+by name, like `--with`.
 """
 
 from __future__ import annotations

package/package.json

@@ -1,7 +1,8 @@
 {
   "name": "@wendellresearch-ai/benchwarmers",
-  "version": "0.2.1",
+  "version": "0.3.0",
   "description": "Self-hosted panel gateway for LLMs: ask a collection of models at once, keep every answer side by side, rotate explorer models in, and let your verdicts pick the next panel.",
+  "mcpName": "io.github.WendellResearch-AI/benchwarmers",
   "license": "MIT",
   "author": "Wendell Research",
   "homepage": "https://github.com/WendellResearch-AI/benchwarmers#readme",

package/perf/run.py

@@ -20,10 +20,14 @@ WHAT IS RECORDED, AND WHY EACH
   queueing, cold start. A large prompt measures THROUGHPUT. A model can be
   fine on one and hopeless on the other, and `nemotron-lightning` (134 s on a
   one-word prompt) is why the distinction earns its place.
-- **Characters per second, not tokens.** The `llm` CLI does not report token
-  counts for these endpoints. Chars/s is a proxy, it is affected by how
-  verbose a model is, and calling it a proxy is more useful than quietly
-  publishing a number that looks like tokens/s and is not.
+- **Characters per second, and tokens per second where they are real.** The
+  `llm` CLI does not report token counts for these endpoints, so chars/s is
+  kept as the proxy every row has; it is affected by how verbose a model is.
+  The built-in client reads the provider's own usage from the stream (Phase
+  2b), and with it `tokens_per_s` is the real thing: output tokens (reasoning
+  included, which is generated like any other) over the time from the first
+  token to the last, so queueing and connection time are not in it. Blank
+  where the provider gave no usage, never estimated.
 - **Outcome counts, split four ways.** answered / refused / error / timeout.
   Collapsing these is the mistake this repo was built around: a safety filter,
   an overloaded service and a slow model are three different facts.
@@ -76,7 +80,9 @@ FIELDS = ["timestamp", "model", "prompt", "n", "p50_s", "p95_s", "min_s",
           # The provider's own words for the worst failure, first 300 chars.
           # Without them `glm-5-3` sat as an unexplained `other` for a month
           # when the message was a plain "404 page not found".
-          "detail"]
+          "detail",
+          # Real output tokens/s from the provider's usage (built-in client only).
+          "tokens_per_s"]
 
 
 def percentile(values: list[float], pct: float) -> float:
@@ -92,14 +98,28 @@ def percentile(values: list[float], pct: float) -> float:
     return ordered[k]
 
 
+def _ask(binary: str, model: str, prompt: str, timeout: int) -> tuple[str, float, float | None]:
+    """ask_once's (text, seconds), plus output tokens/s when the built-in
+    client made the call and the provider reported usage."""
+    import client
+    if client.backend(model) != "builtin":
+        return (*ask_once(binary, model, prompt, timeout), None)
+    res = client.chat(model, prompt, timeout=timeout, purpose="perf")
+    out = (res["usage"] or {}).get("completion_tokens") or (res["usage"] or {}).get("output_tokens")
+    # From the first token on, so connection and queueing time are left out.
+    span = res["seconds"] - (res.get("first_token_s") or 0)
+    rate = out / span if isinstance(out, int) and out > 0 and span > 0 and not res["error"] else None
+    return client.output(res, timeout), res["seconds"], rate
+
+
 def measure(binary: str, model: str, label: str, prompt: str, n: int,
             timeout: int) -> dict:
-    latencies, rates = [], []
+    latencies, rates, token_rates = [], [], []
     counts = {"answered": 0, "refused": 0, "error": 0, "timeout": 0}
     failures: list[str] = []
     details: dict[str, str] = {}
     for i in range(1, n + 1):
-        body, seconds = ask_once(binary, model, prompt, timeout)
+        body, seconds, tokens_per_s = _ask(binary, model, prompt, timeout)
         # A one-word reply is the CORRECT answer to the tiny prompt, so
         # only a genuinely empty one counts as a failure here.
         kind = (outcome(body, min_length=1) or ("answered", ""))[0]
@@ -123,6 +143,8 @@ def measure(binary: str, model: str, label: str, prompt: str, n: int,
             # worse than no number.
             if seconds > 0 and label == "large":
                 rates.append(len(body) / seconds)
+                if tokens_per_s:
+                    token_rates.append(tokens_per_s)
         print(f"    {model} {label} [{i}/{n}] {kind} {seconds:.1f}s", flush=True)
     worst = next((c for c in FAILURE_CLASSES if c in failures), "")
     return {
@@ -139,6 +161,7 @@ def measure(binary: str, model: str, label: str, prompt: str, n: int,
         # is what silently dropped `paywalled` and wrote a blank failure.
         "failure": worst,
         "detail": details.get(worst, ""),
+        "tokens_per_s": round(statistics.median(token_rates), 1) if token_rates else "",
         **counts,
     }
 

package/scripts/check.py

@@ -38,10 +38,12 @@ CHECKS = [
     ("metrics/client.py", ["--selftest"], False),
     ("metrics/bench_collections.py", [], False),
     ("metrics/learn.py", ["--selftest"], False),
+    ("metrics/quota.py", ["--selftest"], False),
     ("metrics/discover.py", ["--selftest"], False),
     ("metrics/test_decide.py", [], False),
     ("metrics/wrong.py", [], False),
     ("metrics/scheduler.py", ["--selftest"], False),
+    ("deploy/container.py", ["--selftest"], False),
     ("scripts/auto_retire.py", [], False),
     ("scripts/onboard.py", ["--selftest"], False),
     ("scripts/export_check.py", ["--selftest"], False),
@@ -52,6 +54,9 @@ CHECKS = [
     ("scripts/test_verdict.py", [], False),
     ("scripts/test_merged.py", [], False),
     ("scripts/test_vision.py", [], False),
+    ("scripts/test_formats.py", [], False),
+    ("scripts/test_route.py", [], False),
+    ("scripts/test_native.py", [], False),
     ("scripts/test_serve_edit.py", [], False),
     ("scripts/test_connect.py", [], False),
     ("scripts/test_auth.py", [], False),
@@ -61,6 +66,8 @@ CHECKS = [
     ("scripts/test_mcp.py", [], False),
     ("scripts/test_cli.py", [], False),
     ("scripts/test_launcher.py", [], False),
+    ("scripts/test_server_json.py", [], False),
+    ("scripts/test_quota.py", [], False),
     ("suite/test_score.py", [], False),
     ("suite/test_mine.py", [], False),
     ("suite/test_seed.py", [], False),

package/scripts/onboard.py

@@ -231,7 +231,10 @@ def main() -> None:
     ap.add_argument("--endpoint", default=None,
                     help="a name from config/endpoints.json; the default endpoint if omitted")
     ap.add_argument("--skip-suite", action="store_true",
-                    help="register and probe only")
+                    help="register and probe only; the scheduled suite passes fill the suite in later")
+    ap.add_argument("--ad-hoc", action="store_true",
+                    help="mark it ad hoc as it is registered: the bench never calls it on its own"
+                         " (for a metered, capped or trial key)")
     args = ap.parse_args()
 
     name = args.served_name
@@ -287,6 +290,12 @@ def main() -> None:
     else:
         register(alias, name, endpoint)
         print(f"  [2/5] registered as {alias}")
+    if args.ad_hoc:
+        # Before the probe, so no scheduled run can pick it up in between.
+        locked.update(DATA / "config" / "model-roles.json",
+                      lambda roles: roles.__setitem__("ad_hoc", sorted({*roles.get("ad_hoc", []), alias})),
+                      ensure_ascii=False)
+        print("        ad hoc: the bench never calls it on its own")
     run("scripts/sync_config.py", timeout=300)
     # Open WebUI's tags name the connection behind the hub; without a hub the
     # endpoint is the connection, and route.connection_of needs no tags.

package/scripts/test_auth.py

@@ -179,7 +179,8 @@ def main() -> int:
             assert call(d, "/api/gateway-key", {"name": "other", "collections": ["nope"]}, **jar)[0] == 400
             assert call(d, "/api/gateway-key", {"name": "other", "rpm": "lots"}, **jar)[0] == 400
             status, _, text = call(g, "/v1/models", **estate)
-            assert [m["id"] for m in json.loads(text)["data"]] == ["bench/vision", "bench/vision:merged"], text
+            assert [m["id"] for m in json.loads(text)["data"]] == ["bench/vision", "bench/vision:merged",
+                                                                   "bench/vision:route"], text
             ask = {"messages": [{"role": "user", "content": "hi"}]}
             status, _, text = call(g, "/v1/chat/completions", {"model": "bench/core", **ask}, **estate)
             assert status == 403 and "may call bench/vision" in text, text