@wendellresearch-ai/benchwarmers 0.6.3 → 0.6.5

package/AGENTS.md

@@ -124,7 +124,9 @@ in `bench.skipped`. One the catalog knows nothing about sits once its model
 popup ticks Vision / OCR.
 After the person checks an answer against the real work, record a verdict
 (dashboard, or `POST /bench/verdict`); verdicts choose the next panels. See
-GATEWAY.md for the whole contract.
+GATEWAY.md for the whole contract. To set an app up on the gateway (opencode,
+VS Code, Codex, Zed, Continue, Cline, Kilo Code, Claude Code, goose, chat
+apps), CLIENTS.md has each one's configuration, file by file.
 
 ## Beyond this machine
 

package/BENCHMARKS.md

@@ -70,7 +70,7 @@ earlier scores on the same tasks, over the last 30 days (or 60).
 
 Whether the provider was there when the bench called it, from the calls the
 bench makes anyway: one cell a day, by provider and then by model, over the
-last 7 days (or 30). At a glance has each model's 7 days too.
+last 7 days (or 30).
 
 - **Down**: the provider was busy or overloaded (a 5xx, high demand, a shared
   pool's 429), unreachable (a broken stream, no reply), timed out (the

package/README.md

@@ -419,6 +419,8 @@ collections and MCP server, with a Copy button: opencode, VS Code chat, Codex
 CLI, Zed, Continue, Cline and Kilo Code, Claude Code, goose, and chat apps
 (LobeHub, AnythingLLM, Cherry Studio, Chatbox, LibreChat). Where keys are
 asked for, it makes one and fills it in, never into a file that may sync.
+The same set-ups are in [CLIENTS.md](CLIENTS.md), with a fresh install's
+address and collections, for reading here or handing to an agent.
 
 <!-- The MCP Registry's proof that the PyPI package is this server (server.json):
 mcp-name: io.github.WendellResearch-AI/benchwarmers -->

package/USAGE.md

@@ -25,15 +25,14 @@ coding agent doing it with you); the gateway's whole contract is in
 
 | Group | Page | What it is for |
 | --- | --- | --- |
-| Observe | **At a glance** | Every model on the bench: whether it works right now, its uptime over 7 days, its role, and its connection. Click a name for its popup: role, alias, facts, options, retire. |
+| Observe | **Benchmarks** | The home page. How models do on the suite and on real work: bugs found against false alarms, who is confidently wrong, each model against its own earlier scores, uptime, speed, and the decision models. Each report shares as an image or its table. |
 | | **Action needed** | What will not fix itself, by the connection it comes from, with Retry and Retire on each model. |
-| | **Benchmarks** | How models do on the suite and on real work: bugs found against false alarms, who is confidently wrong, each model against its own earlier scores, uptime, speed, and the decision models. Each report shares as an image or its table. |
 | | **Quota management** | A tile per provider: what is left, when it resets, and which of its models are resting. Then the week's resets, and what the bench sent. |
 | | **Logs** | Gateway calls (who sat and how it went, never the prompt), changes made from the dashboard, the scheduler, and onboarding runs. |
 | Gateway | **Endpoints** | Where models are served from. Add an endpoint (a provider's sign-in, or any OpenAI-compatible API key), change a key, and see what each one serves. A model server running on this machine, at its usual port or any other its process is listening on (LM Studio, Jan, llama.cpp, vLLM and the like), is offered under Found on this machine: Add opens the dialog with it filled in, to test and save. Also your own app registrations and rate limits per connection. |
-| | **Models** | Every model on the bench, by the connection it comes from. |
+| | **Models** | Every model on the bench, by the connection it comes from: what it holds, its reply time and its limits. One search above every table, each table folds, and Collapse all / Expand all. Click a name for its popup: role, alias, facts, options, retire. |
 | | **Collections** | The named panels your apps ask for, and the gateway's settings. |
-| | **Clients** | Each popular app's set-up (opencode, VS Code, Codex, Zed, Continue, Cline, Kilo Code, Claude Code, goose, chat apps), filled in from this install, with Copy and a gateway key made on the spot. |
+| | **Clients** | Each popular app's set-up (opencode, VS Code, Codex, Zed, Continue, Cline, Kilo Code, Claude Code, goose, chat apps), filled in from this install, with Copy and a gateway key made on the spot. [CLIENTS.md](CLIENTS.md) has the same, for a fresh install. |
 | | **Discover** | Models worth adding, from public catalogs, with the ways this install can reach each one. |
 | Control | **Settings** | Theme, gateway keys, the scheduler, the model client, providers whose free models work only in their own app, export and import, how the install is reached (shown only), and Restart. |
 | | **Info** | The source repository, the version (linking its release), Python, data folder and counts. |
@@ -52,8 +51,7 @@ clears the alert) and Retire. A group of two or more has Retry all and Retire
 all; a connection held because every model on it fails says so above its
 models, with any link the provider gave. Automatic retirements from the last
 week are folded at the end, each with Restore. Its count sits beside it in the
-sidebar, and At a glance opens on one line saying what is waiting and how
-soon. An alert whose models have all been retired goes at once. A model that
+sidebar. An alert whose models have all been retired goes at once. A model that
 is out of quota is not counting down: days the bench held its calls back until
 a stated reset, an account's shared free quota was spent, or the provider said
 the model's free quota is used up, pause its clock, and while it is still out

package/VERSION

@@ -1 +1 @@
-0.6.3
+0.6.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wendellresearch-ai/benchwarmers",
-  "version": "0.6.3",
+  "version": "0.6.5",
   "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",

package/scripts/check.py

@@ -49,6 +49,7 @@ CHECKS = [
     ("scripts/onboard.py", ["--selftest"], False),
     ("scripts/export_check.py", ["--selftest"], False),
     ("scripts/notice.py", ["--check"], False),
+    ("scripts/clients_doc.py", ["--check"], False),       # CLIENTS.md is what the Clients page shows
     ("scripts/check_themes.py", [], False),
     ("scripts/test_owui.py", [], False),
     ("scripts/test_hubs.py", [], False),

package/scripts/clients_doc.py

@@ -0,0 +1,98 @@
+"""Write CLIENTS.md from the dashboard's Clients page, so each app's set-up is in
+the repository for people and agents to read, and never drifts from the page.
+
+    python scripts/clients_doc.py           # write CLIENTS.md
+    python scripts/clients_doc.py --check   # fail if CLIENTS.md is not what it would write (check.py)
+
+One source: serve.client_setups, the function the Clients page draws. The page
+fills in this install (its address, its collections, a key made on the spot);
+here they are the defaults a fresh install has: the server at 127.0.0.1:8122,
+collections `core` and `fast`, and every call needing a gateway key, read from
+BENCHWARMERS_KEY. The page's HTML (code, links) becomes Markdown.
+"""
+
+from __future__ import annotations
+
+import html
+import os
+import pathlib
+import re
+import sys
+import tempfile
+
+ROOT = pathlib.Path(__file__).resolve().parent.parent
+OUT = ROOT / "CLIENTS.md"
+COLLECTIONS = [("core", None, False), ("fast", None, False)]
+
+
+def md(text: str) -> str:
+    """The page's inline HTML as Markdown: code, links, line breaks; any other tag dropped."""
+    text = re.sub(r"<code>(.*?)</code>", lambda m: "`" + html.unescape(m.group(1)) + "`", text)
+    text = re.sub(r"<a [^>]*href='([^']*)'[^>]*>(.*?)</a>", r"[\2](\1)", text)
+    text = re.sub(r"<br\s*/?>", "  \n", text)
+    return html.unescape(re.sub(r"<[^>]+>", "", text))
+
+
+def fence(lang: str, text: str) -> str:
+    tick = "````" if "```" in text else "```"
+    return f"{tick}{lang}\n{text}\n{tick}"
+
+
+def render() -> str:
+    # serve reads the data folder at import: an empty temp one, never the install's.
+    os.environ.setdefault("BENCHWARMERS_HOME", tempfile.mkdtemp(prefix="bench-clients-doc-"))
+    sys.path[:0] = [str(ROOT), str(ROOT / "metrics"), str(ROOT / "suite")]
+    import serve
+    root = serve.DEFAULT_SERVER
+    out = ["# Setting up clients",
+           "",
+           "Every app Benchwarmers has a ready set-up for, as the dashboard's **Clients** page shows it,"
+           " written out for people and for agents setting an app up. The page fills in your own install:"
+           " its address, your collections and a gateway key made on the spot. Here they are a fresh"
+           f" install's: the server at `{root}`, the collections `core` and `fast`, and a gateway key"
+           f" read from `{serve.KEY_ENV}`. Replace them with yours.",
+           "",
+           "Agents get a collection in route mode (one member with its tools, `bench/<collection>:route`);"
+           " chat windows a merged answer (`:merged`); the MCP server the panel itself, as ask and verdict"
+           f" tools. Gateway `{root}/v1` · MCP server `{root}/mcp`.",
+           "",
+           f"_Generated from `serve.client_setups` by `scripts/clients_doc.py`; each app's format follows its"
+           f" own docs as read on {serve.CLIENTS_READ}. Do not edit by hand: run the script._",
+           "",
+           "## The gateway key",
+           "",
+           "When every call needs a key (`BENCHWARMERS_REQUIRE_KEYS=1`, or an admin token), make one on"
+           " Settings or the Clients page. Most set-ups read it from"
+           f" `{serve.KEY_ENV}`:",
+           ""]
+    for s in serve._env_snips(serve.KEY_ENV):
+        out += [f"**{s['label']}**", "", fence(s["lang"], s["text"]), ""]
+    setups = serve.client_setups(root, COLLECTIONS, True)
+    out += ["## Contents", ""] + [f"- [{c['name']}](#{c['id']})" for c in setups] + [""]
+    for c in setups:
+        out += [f"<a id='{c['id']}'></a>", "", f"## {c['name']}", "", md(c["gets"]), "",
+                f"**Where it goes:** {md(c['where'])}", ""]
+        for s in c["blocks"]:
+            out += [f"**{s['label']}**", "", fence(s["lang"], s["text"]), ""]
+        docs = ", ".join(f"[{t}]({u})" for u, t in c["docs"])
+        out += [f"**Watch for:** {md(c['watch'])}", "",
+                f"Written from {docs}, read {serve.CLIENTS_READ}"
+                + (f": {md(c['docs_note'])}" if c.get("docs_note") else "."), ""]
+    return "\n".join(out).rstrip() + "\n"
+
+
+def main() -> int:
+    text = render()
+    if "--check" in sys.argv:
+        if not OUT.exists() or OUT.read_text(encoding="utf-8") != text:
+            print("CLIENTS.md is not what serve.client_setups gives: run python scripts/clients_doc.py")
+            return 1
+        print("CLIENTS.md matches the Clients page")
+        return 0
+    OUT.write_text(text, encoding="utf-8", newline="\n")
+    print(f"wrote {OUT.name}")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/serve.py

@@ -2254,27 +2254,6 @@ def open_count(groups: dict) -> int:
     return sum(len(g["rows"]) + len(g["held"]) for g in groups.values())
 
 
-def and_list(parts: list[str]) -> str:
-    return " and ".join(parts) if len(parts) < 3 else ", ".join(parts[:-1]) + " and " + parts[-1]
-
-
-def actions_line(groups: dict) -> str:
-    """At a glance's one line about Action needed, when anything is open:
-    what is waiting and how soon, and the way to its page."""
-    rows = [r for g in groups.values() for r in g["rows"]]
-    held = sum(len(g["held"]) for g in groups.values())
-    if not rows and not held:
-        return ""
-    models = sum(1 for r in rows if r["alias"])
-    what = and_list([f"{n} {word}{'s' * (n != 1)}" for n, word in
-                     ((models, "model"), (held, "connection"), (len(rows) - models, "notice")) if n])
-    soon = min(r["days"] for r in rows) if rows else 99
-    many = sum(r["days"] == soon for r in rows)
-    when = (f"; {many} retire{'s' * (many == 1)} in {soon} day{'s' * (soon != 1)}" if soon < 99 else "")
-    return (f"<p class='a-line'><strong>Action needed</strong> <span>{what}{when}.</span>"
-            " <button type='button' class='linkish' data-go='needs'>Review →</button></p>")
-
-
 def actions_page(groups: dict, auto: list) -> str:
     """Action needed, its own page since 2026-11-02: it sat above every page
     and grew to a screen and more of red. Grouped by connection, because
@@ -2373,167 +2352,6 @@ def price_of(alias: str, served: str, f: dict) -> str:
     return "paid" if sources.is_free(f.get("cost")) is False else "unknown"
 
 
-def facet(name: str, label: str, values) -> str:
-    """A filter by one column's value, beside a panel's search box (bench.js
-    matches it against each row's data-<name>)."""
-    opts = "".join(f"<option value='{esc(v)}'>{esc(v)}</option>" for v in sorted(set(values) - {""}))
-    return (f"<label>{esc(label)} <select data-facet='{esc(name)}'><option value=''>any</option>"
-            f"{opts}</select></label>")
-
-
-def facts_cells(alias: str, f: dict | None = None) -> str:
-    """Context window and capabilities, from metrics/catalog.py; a value you
-    entered is marked, and an unknown one says so rather than looking empty."""
-    f = f or catalog.info(alias)
-    ctx = f.get("context_window")
-    src = f["sources"].get("context_window", "")
-    ctx_cell = (f"<td class='num' data-sort='{ctx}' title='{esc(src)}'>"
-                f"{ctx // 1000:,}k{'*' if src == 'you' else ''}</td>" if ctx
-                else "<td class='num muted' data-sort='-1' title='unknown: fill it in the model popup'>?</td>")
-    caps = [CAP_LABELS[c] for c in catalog.CAPS if f.get(c)]
-    return ctx_cell + (f"<td class='caps'>{' '.join(chip(c, '') for c in caps)}</td>" if caps
-                       else "<td class='muted'>—</td>")
-
-
-def roster() -> str:
-    """Every model on one table: its state, what it has done, what it scored.
-
-    Replaces a grid of offload cards that answered only "what was it used
-    for". Status first, because the first question on opening this page is
-    whether anything is broken; real work second; benchmarks after."""
-    try:
-        from offload import summary
-        work = summary()
-    except Exception:                              # noqa: BLE001
-        work = {}
-    try:
-        import auto_retire
-        rows, answers, _ = auto_retire.evidence()
-    except Exception:                              # noqa: BLE001
-        rows, answers = [], {}
-    last = dict(answers)
-    for r in rows:
-        if r["answered"]:
-            last[r["alias"]] = max(last.get(r["alias"], r["date"]), r["date"])
-    suite: dict[str, list[int]] = {}
-    for (_task, model), hit in suite_scores().items():
-        if (m := re.match(r"(\d+)/(\d+)", hit["found"])):
-            got = suite.setdefault(model, [0, 0, 0])
-            got[0] += int(m.group(1)); got[1] += int(m.group(2)); got[2] += 1
-    tiny = {r["model"]: float(r["p50_s"]) for r in perf_rows()
-            if r["prompt"] == "tiny" and int(r.get("answered") or 0) > 0}
-    status = model_status()
-    new = recently_onboarded()
-    stuck = onboard_issues(status)
-    directory = model_directory()
-    severity = {"err": 0, "warn": 1, "": 2, "ok": 3}
-
-    up = uptime_rows()["models"]
-    body = ["<table><thead><tr>", th("Model"), th("Status"), th("Uptime 7d", True), th("Role"),
-            th("Context", True), th("Can"),
-            th("Offloads", True), th("Helped"), th("Suite", True),
-            th("Reply time", True), th("Last answer"), "</tr></thead><tbody>"]
-    # The model browser: each row says its connection, status, role and price
-    # for the toolbar's filters, and the filters offer what the rows hold.
-    offered: dict[str, list[str]] = {"connection": [], "status": [], "role": [], "price": []}
-
-    def browse(**values) -> str:
-        for k, v in values.items():
-            offered[k].append(v)
-        return "".join(f" data-{k}='{esc(v)}'" for k, v in values.items())
-    for alias in sorted(status, key=lambda a: (-work.get(a, {}).get("offloads", 0), a)):
-        label, css, why = status[alias]
-        f = catalog.info(alias)
-        served = directory.get(alias, {}).get("served") or retired().get(alias, {}).get("served", "")
-        w = work.get(alias, {})
-        judged, useful = w.get("judged", 0), w.get("useful", 0)
-        if judged:
-            rate = useful / judged
-            helped = (f"<span class='helped'><span class='meter'><span"
-                      f"{' class=low' if rate < .5 else ''} style='width:{max(4, round(rate * 100))}%'>"
-                      f"</span></span><span class='mono'>{useful}/{judged}</span></span>")
-        else:
-            rate, helped = -1, "<span class='muted'>not judged</span>"
-        s = suite.get(alias)
-        suite_cell = (f"<td class='num' data-sort='{s[0] / s[1]:.3f}'"
-                      f" title='{s[2]} task{'s' * (s[2] != 1)}'>{s[0]}/{s[1]}</td>"
-                      if s and s[1] else "<td class='num muted' data-sort='-1'>—</td>")
-        p50 = tiny.get(alias)
-        seen = last.get(alias)
-        role = directory.get(alias, {}).get("role") or retired().get(alias, {}).get("role", "")
-        conn = catalog.connection_of(alias)
-        if conn == "?":                    # retired: no registration left to say, but its record does
-            conn = retired().get(alias, {}).get("tag", "").split("/")[0] or "unknown"
-        attrs = browse(connection=conn, status=re.sub(r" in \d+d$", "", label),
-                       role=role, price=price_of(alias, served, f))
-        body.append(
-            f"{tr(alias, attrs)}<td>{named(alias)}</td>"
-            f"<td data-sort='{severity[css]}'>{chip(label, css, why)}"
-            + (" " + chip("new", "new", "onboarded this week") if alias in new else "")
-            + (" " + chip("onboard stopped", stuck[alias]["css"],
-                          f"onboarding stopped at the {stuck[alias].get('step', 'probe')}:"
-                          f" {stuck[alias]['label']}. Open the model for the error.")
-               if alias in stuck else "")
-            + "</td>" + uptime_cell(up.get(alias)) + f"<td>{esc(role)}</td>{facts_cells(alias, f)}"
-            f"<td class='num'>{w.get('offloads', 0)}</td>"
-            f"<td data-sort='{rate:.3f}'>{helped}</td>{suite_cell}"
-            + (f"<td class='num' data-sort='{p50}'>{p50:.1f}s</td>" if p50 is not None
-               else "<td class='num muted' data-sort='99999'>—</td>")
-            + (f"<td class='mono' data-sort='{seen}'>{seen}</td>" if seen
-               else "<td class='muted' data-sort='0'>never</td>")
-            + "</tr>")
-    # Decision models are not in the llm yaml, so no probe, suite or reply time
-    # applies; the role column says what they are and what they answer.
-    for alias, d in sorted(decide.load().items()):
-        w = work.get(alias, {})
-        body.append(
-            "<tr" + browse(connection=d.get("via", ""), status="no probe", role="decision",
-                           price="free" if ":free" in d["served"] else "unknown") + ">"
-            f"<td>{named(alias)}</td>"
-            "<td data-sort='2'><span class='muted' title='decision models have no"
-            " availability probe yet'>—</span>"
-            + (" " + chip("new", "new", "onboarded this week") if d.get("onboarded", "") >=
-               (datetime.date.today() - datetime.timedelta(days=7)).isoformat() else "")
-            + "</td>" + uptime_cell(up.get(alias)) + f"<td>decision · {esc(', '.join(d['question_types']))}</td>"
-            "<td class='num muted' data-sort='-1'>—</td><td class='muted'>—</td>"
-            f"<td class='num'>{w.get('offloads', 0)}</td>"
-            "<td data-sort='-1'><span class='muted'>not judged</span></td>"
-            "<td class='num muted' data-sort='-1'>—</td><td class='num muted' data-sort='99999'>—</td>"
-            "<td class='muted' data-sort='0'>—</td></tr>")
-    body.append("</tbody></table>")
-    key = """<dl>
-      <dt>Status</dt><dd>From the latest probe. <em>healthy</em> answered;
-        <em>throttled</em> and <em>slow</em> are rate limits and timeouts, which
-        never retire a model inside 7 days; <em>retiring in Nd</em> is the
-        countdown before the bench retires it on its own. Hover a chip for the
-        provider's own words.</dd>
-      <dt>Uptime 7d</dt><dd>Of its calls in the last 7 days that say anything about the provider,
-        the share answered: busy, unreachable and timed out count against it; our own quota,
-        pacing and setup problems are left out. Benchmarks has it a day at a time.</dd>
-      <dt>Offloads, Helped</dt><dd>Real work handed over, and how much of what
-        was judged actually helped. Benchmarks never count here.</dd>
-      <dt>Suite</dt><dd>Verified findings located across the internal tasks:
-        the tier that decides routing.</dd>
-      <dt>Reply time</dt><dd>Median seconds to answer a one-word prompt: the
-        serving path's overhead, not the model's worth.</dd>
-      <dt>Last answer</dt><dd>The latest real answer from any source: probe,
-        suite or a useful offload. Seven days without one retires a model.</dd>
-      <dt>decision</dt><dd>A decision model: it answers typed questions (yes/no,
-        choice, score) with probabilities through a Decisions API, never chat.
-        The role column lists the question types it takes. No probe, suite or
-        reply time applies to it.</dd>
-      <dt>Filters</dt><dd>Connection is where a model's calls go (its Open WebUI
-        tag, or its own endpoint). Price: <em>free</em> for a <code>:free</code>
-        slug, a local model, a list price of 0 or a free-tier listing;
-        <em>paid</em> for a subscription or a list price; else unknown. Combine
-        them with the search box.</dd>
-    </dl>"""
-    facets = "".join(facet(k, k.title(), v) for k, v in offered.items())
-    return card("At a glance", "Every model on the bench: whether it works right "
-                "now, what it has been used for, and what it scored.",
-                "".join(body), key, facets)
-
-
 def on_bench() -> set[str]:
     """Aliases registered now: what the charts draw."""
     return set(registered().values())
@@ -2883,7 +2701,14 @@ def vision_card() -> str:
         come from these alone. <b>Useful, no claims</b> is a verdict without them, which on thicken's
         photos means the seat agreed with another seat. Agreement is not a check, since seats fail
         together, so it is never counted as right.</dd>
-      <dt>The number beside a name</dt><dd>How much there is: fields asked on the tasks, seats on real work.</dd>
+      <dt>The real-work chart</dt><dd>Read as Real work is. The dot is the share of its judged seats your
+        verdict called useful (checked and helped, or useful with no claims), bigger for more judged; the
+        line is where that share likely sits (95%), and the order is by its low end, so a lot of evidence
+        outranks a little. Green when even the low end is over half, red when even the high end is under,
+        grey when there are too few to say. Beside it: useful of judged, how often it answered at all, and
+        the claims checked by hand, the only count of right and wrong. Seats not judged or not answered
+        are in those words and the table, not the dot.</dd>
+      <dt>The number beside a name</dt><dd>How much there is: fields asked on the tasks.</dd>
     </dl>"""
     held = (f"<p class='hint'>The vision tasks wait until {esc(v['held'])}: no run picks them before then"
             " (<code>vision_from</code> in <code>config/suite.json</code>; remove it to start them now).</p>"
@@ -2891,7 +2716,7 @@ def vision_card() -> str:
     return (f"<div class='panel'>{SHARE}<h2>Vision</h2><p class='sub'>Reading images: fields read right on pictures with"
             " a known answer, then real work on images and how much of it was actually checked.</p>" + held +
             "<h3 class='chart-head'>On the vision tasks, field by field</h3><div class='chart' id='chart-vision-fields'></div>"
-            "<h3 class='chart-head'>On real work with images, by how each seat was judged</h3>"
+            "<h3 class='chart-head'>On real work with images: judged useful, and how sure that is</h3>"
             "<div class='chart' id='chart-vision-work'></div>"
             "<details class='key'><summary>Table</summary><div class='toolbar'><input type='search'"
             " data-filter placeholder='Filter by model or connection' aria-label='Filter rows'>"
@@ -3378,15 +3203,6 @@ def _uptime_window(cells: list, days: int) -> dict:
             "median": seconds[len(seconds) // 2] if seconds else None}
 
 
-def uptime_cell(cells: list | None, days: int = UPTIME_DAYS[0]) -> str:
-    """At a glance's Uptime column: answered of counted over the last `days`."""
-    t = _uptime_window(cells or [], days)
-    if t["pct"] is None:
-        return "<td class='num muted' data-sort='-1' title='no counted calls'>—</td>"
-    return (f"<td class='num{' err' if t['pct'] < .8 else ''}' data-sort='{t['pct']:.4f}'"
-            f" title='{t['up']} of {t['up'] + t['down']} answered'>{t['pct'] * 100:.1f}%</td>")
-
-
 def _why(kinds: dict, which) -> str:
     return ", ".join(f"{k} {n}" for k, n in sorted(kinds.items(), key=lambda kv: -kv[1]) if which(k))
 
@@ -3415,7 +3231,7 @@ def uptime_data() -> dict:
 
 def uptime_card() -> str:
     """Was the provider there when the bench called it: a cell a day, by
-    provider then by model; 7 days, or 30 on choice (bench.js swaps them)."""
+    provider or by model, as its By says; 7 days, or 30 on choice (bench.js swaps them)."""
     u = uptime_rows()
     first, gone, names = UPTIME_DAYS[0], retired(), {}
 
@@ -3489,9 +3305,11 @@ def uptime_card() -> str:
               "<span><i style='background:var(--signal)'></i>some down, 80% or more answered</span>"
               "<span><i style='background:var(--err)'></i>under 80% answered</span>"
               "<span><i style='background:none;border:1px solid var(--line)'></i>no counted calls</span></div>")
+    # The window and one view at a time: by provider (shorter, the default) or by model (bench.js).
     pick = ("<div class='toolbar'><label>Days <select data-uptime-days>"
             + "".join(f"<option value='{w}'>last {w}</option>" for w in UPTIME_DAYS)
-            + f"</select></label></div>{legend}")
+            + "</select></label><label>By <select data-uptime-by><option value='providers'>provider</option>"
+            f"<option value='models'>model</option></select></label></div>{legend}")
     key = """<dl>
       <dt>Down</dt><dd>The provider was busy or overloaded (a 5xx, high demand, a shared pool's
         429), unreachable (a broken stream, no reply), timed out (the gateway's 420 s included), or
@@ -3517,9 +3335,7 @@ def uptime_card() -> str:
                       " and every real call, one cell a day. Busy, unreachable and timed out count as down;"
                       " our own quota, our own pacing and setup problems are left out.",
                       "chart-uptime-models", "".join(window(w) for w in UPTIME_DAYS), key,
-                      extra=pick + "<h3 class='chart-head'>By provider</h3>"
-                                   "<div class='chart' id='chart-uptime-providers'></div>"
-                                   "<h3 class='chart-head'>By model</h3>",
+                      extra=pick + "<div class='chart' id='chart-uptime-providers'></div>",
                       after="".join(odd(w) for w in UPTIME_DAYS))
 
 
@@ -3846,10 +3662,20 @@ def wrong_card() -> str:
 
 def benchmarks_tab() -> str:
     data = json.dumps(charts_data(), ensure_ascii=False).replace("</", "<\\/")
-    # Trend, Uptime and Performance side by side: what moves over time, in one stretch.
-    return (overview_card() + bughunt_card() + work_card() + wrong_card() + suite_card() + vision_card()
-            + trend_card() + uptime_card() + perf_card() + decision_card() + external_card()
-            + f"<script type='application/json' id='charts-data'>{data}</script>")
+    # Who to trust, then what moves day to day (Uptime, Trend, Performance), then Real work,
+    # then the rest in the order they had. Each report gets an id; a bar of links to them
+    # stays at the top while the page scrolls, the one in view marked (bench.js jumpSpy).
+    reports = (("trust", "Who to trust", overview_card), ("uptime", "Uptime", uptime_card),
+               ("trend", "Trend", trend_card), ("performance", "Performance", perf_card),
+               ("work", "Real work", work_card), ("hunt", "Bug hunt", bughunt_card),
+               ("wrong", "Confidently wrong", wrong_card), ("code", "Our code", suite_card),
+               ("vision", "Vision", vision_card), ("decisions", "Decision models", decision_card),
+               ("external", "External", external_card))
+    links = "".join(f"<a href='#report-{key}'>{esc(label)}</a>" for key, label, _ in reports)
+    cards = "".join(card().replace("<div class='panel'", f"<div class='panel report' id='report-{key}'", 1)
+                    for key, _, card in reports)
+    return (f"<nav class='jump' aria-label='On this page'><p class='jump-title'>On this page</p>{links}</nav>{cards}"
+            f"<script type='application/json' id='charts-data'>{data}</script>")
 
 
 def perf_card() -> str:
@@ -3925,12 +3751,12 @@ def favicon(theme: str) -> str:
     return f"<link rel='icon' type='image/svg+xml' href='data:image/svg+xml,{quote(svg)}'>"
 
 
-def render(theme: str | None = None, shown: str = "glance") -> str:
+def render(theme: str | None = None, shown: str = "suite") -> str:
     with locked.snapshot():                    # each file read once a render (metrics/locked.py)
         return _page(theme, shown)
 
 
-def _page(theme: str | None, shown: str = "glance") -> str:
+def _page(theme: str | None, shown: str = "suite") -> str:
     theme = theme if theme in themes() else saved_theme()      # a ?theme preview, else the saved one
     return (
         f"<!doctype html><html lang='en' data-theme='{esc(theme)}'><meta charset='utf-8'>"
@@ -4406,9 +4232,10 @@ def collections_tab() -> str:
 # A grouped sidebar since 2026-10-19 (RELEASE-PLAN "The dashboard"): eight
 # pages were too many for a row of tabs. Collections holds the gateway's
 # requests under the collections they ran; Benchmarks holds Performance.
-SECTIONS = (("Observe", (("glance", "At a glance", "roster"),
+# At a glance (every model's state and uptime) was retired 2026-11-10: Benchmarks,
+# Uptime and Models said it all again. Benchmarks is the home page.
+SECTIONS = (("Observe", (("suite", "Benchmarks", "benchmarks_tab"),
                          ("needs", "Action needed", "actions_page"),
-                         ("suite", "Benchmarks", "benchmarks_tab"),
                          ("quotas", "Quota management", "quotas_page"),
                          ("logs", "Logs", "logs_page"))),
             ("Gateway", (("endpoints", "Endpoints", "endpoints_page"),
@@ -4433,7 +4260,9 @@ def collections_and_gateway() -> str:
 # names an environment variable, or the app asks, or (an MCP header in Zed,
 # Cline or goose, which take no variable) KEY_HOLDER stays for the reader.
 # Only blocks marked `key` (shell lines, fields typed into an app) get the key
-# made on the page, swapped in for KEY_HOLDER by bench.js.
+# made on the page, swapped in for KEY_HOLDER by bench.js. CLIENTS.md is written
+# from client_setups (scripts/clients_doc.py): after a change here, run it, as
+# check.py fails while the file differs.
 KEY_HOLDER = "<your gateway key>"
 KEY_ENV = "BENCHWARMERS_KEY"         # the plugin's and the CLI's name for it too
 CLIENTS_READ = "2026-11-08"
@@ -4729,20 +4558,20 @@ def _snip_html(group: str, i: int, s: dict) -> str:
 def pane(key: str, actions: tuple | None = None) -> str:
     """One tab's page, as its section holds it. The page renders the tab shown
     (tabs); bench.js fetches the others from GET /api/pane when first opened."""
-    if key in ("needs", "glance"):
+    if key == "needs":
         groups, auto = actions or open_actions()
-        return actions_page(groups, auto) if key == "needs" else actions_line(groups) + roster()
+        return actions_page(groups, auto)
     return globals()[next(fn for k, _, fn in TABS if k == key)]()
 
 
-def tabs(shown: str = "glance") -> str:
-    # Read once a render: Action needed's count in the sidebar, and its page or
-    # At a glance's line about it when one of those is shown. Only the tab
-    # shown is rendered (the bench_tab cookie, set by bench.js): rendering every
-    # one took seconds. The rest are empty until opened (data-lazy).
+def tabs(shown: str = "suite") -> str:
+    # Read once a render: Action needed's count in the sidebar, and its page when
+    # shown. Only the tab shown is rendered (the bench_tab cookie, set by bench.js):
+    # rendering every one took seconds. The rest are empty until opened (data-lazy).
+    # A tab that is gone (a cookie still saying glance) shows the home page.
     actions = open_actions()
     count = {"needs": open_count(actions[0])}
-    shown = shown if any(k == shown for k, _, _ in TABS) else "glance"
+    shown = shown if any(k == shown for k, _, _ in TABS) else "suite"
 
     def button(key: str, label: str) -> str:
         n = count.get(key)
@@ -5438,8 +5267,17 @@ def models_page() -> str:
     return ("<div class='panel'><h2>Models</h2><p class='sub'>Every model on the bench by the connection it"
             " comes from, with the endpoint it is reached through and what is known of its limits. A name"
             " opens the model.</p><div class='toolbar'><input type='search' data-filter='page'"
-            " placeholder='Find a model or connection' aria-label='Filter models'></div></div>"
-            + "".join(cards))
+            " placeholder='Find a model or connection' aria-label='Filter models'>"
+            f"{fold_toggle(True)}</div></div>" + "".join(cards))
+
+
+def fold_toggle(all_open: bool) -> str:
+    """One labelled toggle over a page's folds (Models' connections, Logs' logs), as
+    accordions do it (GOV.UK's "Show all sections"): Collapse all while every one is
+    open, else Expand all. bench.js foldButtons keeps it right as folds change."""
+    return (f"<button class='fold-all' type='button' data-fold-all='{'close' if all_open else 'open'}'"
+            f" aria-expanded='{str(all_open).lower()}'><span class='fold-chev' aria-hidden='true'></span>"
+            f"{'Collapse all' if all_open else 'Expand all'}</button>")
 
 
 def quota_tiles() -> list[dict]:
@@ -5615,11 +5453,25 @@ def logs_page() -> str:
                       for p in sorted((DATA / "runs").glob("onboard-*.log"), reverse=True)[:30])
 
     def section(title, body, n, open_=False):
-        return (f"<details class='panel'{' open' if open_ else ''}><summary><b>{esc(title)}</b>"
+        return (f"<details class='panel'{' open' if open_ else ''} data-log='{esc(title)}'><summary><b>{esc(title)}</b>"
                 f" <span class='muted'>· {n}</span></summary>{body}</details>")
+    # Export: a popup for the format and the logs (bench.js), each as the page shows it,
+    # the filter applied. Text, every log chosen in one file to read; CSV, a file per log.
+    picks = "".join(f"<label class='check'><input type='checkbox' name='lx-log' value='{esc(t)}' checked> {esc(t)}</label>"
+                    for t in ("Gateway calls", "Changes", "Scheduler", "Onboarding runs"))
+    export = ("<dialog id='logs-export-dialog' aria-labelledby='lx-title'><h2 id='lx-title'>Export logs</h2>"
+              "<fieldset><legend class='field'>Format</legend>"
+              "<label class='check'><input type='radio' name='lx-format' value='txt' checked> Text (.txt)</label>"
+              "<label class='check'><input type='radio' name='lx-format' value='csv'> CSV (.csv)</label></fieldset>"
+              f"<fieldset><legend class='field'>Logs</legend>{picks}</fieldset>"
+              "<p class='rules' id='lx-note'></p><div class='errors' id='lx-errors' role='alert'></div>"
+              "<div class='actions'><button class='btn' type='button' id='lx-cancel'>Cancel</button>"
+              "<button class='btn primary' type='button' id='lx-go'>Export</button></div></dialog>")
     return (
-        "<div class='panel'><input data-filter='page' placeholder='Filter every log on this page'"
-        " aria-label='Filter logs'></div>"
+        "<div class='panel'><div class='toolbar'><input data-filter='page' placeholder='Filter every log on this page'"
+        " aria-label='Filter logs'>"
+        "<button class='btn small' type='button' id='logs-export'>Export…</button>"
+        f"{fold_toggle(False)}</div></div>{export}"
         + section("Gateway calls", "<table><thead><tr>" + th("When") + th("Collection") + th("Shape")
                   + th("Seats") + f"</tr></thead><tbody>{''.join(gw)}</tbody></table>", len(gw), True)
         + section("Changes", "<table><thead><tr>" + th("When") + th("By") + th("Model") + th("What")
@@ -5938,8 +5790,7 @@ class Handler(gateway.Handler):
         if self.path == "/api/needs":                  # after an action there: that page, not the whole one
             with locked.snapshot():
                 groups, auto = open_actions()
-                self._json(200, {"html": actions_page(groups, auto), "line": actions_line(groups),
-                                 "count": open_count(groups)})
+                self._json(200, {"html": actions_page(groups, auto), "count": open_count(groups)})
             return
         if path == "/api/pane":                        # a tab's page, when first opened (bench.js load)
             key = q.get("tab", [""])[0]
@@ -6058,7 +5909,7 @@ class Handler(gateway.Handler):
         theme = q.get("theme", [None])[0]               # a preview; else the saved one
         shown = re.search(r"(?:^|;)\s*bench_tab=(\w+)", self.headers.get("Cookie") or "")   # the tab last shown
         try:
-            page = render(theme, shown.group(1) if shown else "glance")
+            page = render(theme, shown.group(1) if shown else "suite")
         except RuntimeError as exc:                    # endpoints.json unreadable: say which and why
             page = f"<!doctype html><meta charset='utf-8'><title>Benchwarmers</title><p>{esc(exc)}</p>"
         self._page(page)                               # gzipped beyond this machine (_length)

package/ui/bench.css

@@ -117,8 +117,6 @@ main { max-width: var(--page, 104rem); margin: 0 auto; padding: 0 var(--gutter)
 
 /* Action needed: its own page, a panel per connection as on Models, each a
    table (name, state, the provider's words, actions). */
-.a-line { display: flex; flex-wrap: wrap; align-items: baseline; gap: .25rem .6rem; margin: 1.4rem 1.6rem 0;
-          padding: .45rem .75rem; background: var(--err-ground); border-radius: var(--radius); }
 .a-head { display: flex; flex-wrap: wrap; align-items: baseline; gap: .25rem .6rem; }
 .a-head .why { font-size: .875rem; }
 .a-head .why:empty { display: none; }
@@ -359,8 +357,7 @@ dialog h2 { margin: 0 0 .2rem; font-size: 1.375rem; font-weight: 700; }
 :is(dialog, .form-panel) .members .m-row { border-bottom: 1px solid var(--line); }
 :is(dialog, .form-panel) .members .m-row:nth-child(odd) { background: color-mix(in oklab, var(--band) 45%, transparent); }
 :is(dialog, .form-panel) .members .m-row:has(input:checked) { box-shadow: inset 3px 0 0 var(--bench-seat, var(--brand)); }
-:is(dialog, .form-panel) .members .m-name a.model { color: var(--ink); text-decoration: none; font-size: .875rem; }
-:is(dialog, .form-panel) .members .m-name a.model:hover { text-decoration: underline; }
+:is(dialog, .form-panel) .members .m-name a.model { font-size: .875rem; }
 :is(dialog, .form-panel) .members .m-ctx { font-family: var(--font-mono); font-size: .8125rem; color: var(--muted); text-align: right; }
 :is(dialog, .form-panel) .members .m-can { display: flex; flex-wrap: wrap; gap: .2rem; }
 :is(dialog, .form-panel) .members .m-can .chip { font-size: .6875rem; padding: 0 .4rem; }
@@ -539,8 +536,44 @@ pre.log.answer { max-height: 40vh; }
 .chart.muted { font-size: .875rem; }
 h3.chart-head { font-size: 1rem; margin: 1rem 0 .2rem; }
 .chart-label { font-size: .875rem; font-weight: 600; margin: .9rem 0 .1rem; }
+/* Benchmarks' "On this page": a link per report, the one in view marked (bench.js jumpSpy),
+   after the U.S. Web Design System's in-page navigation. Dressed as the sidebar is: the
+   theme's frosted rail, muted links, the highlighter stroke on the current one.
+   Narrow: a bar floating at the top of the page, sideways when it does not fit.
+   Wide: a rail beside the reports, sticky, every link on its own line. */
+.jump { position: sticky; top: .75rem; z-index: 4; display: flex; align-items: center; gap: 1rem;
+        margin: .9rem 1.2rem 0; padding: .45rem 1.1rem; overflow-x: auto; scrollbar-width: none;
+        background: var(--rail, var(--paper-2)); border: 1px solid var(--line); border-radius: 999px;
+        -webkit-backdrop-filter: var(--glass, blur(10px)); backdrop-filter: var(--glass, blur(10px));
+        box-shadow: 0 10px 28px -16px rgb(0 0 0 / .4); }
+.jump::-webkit-scrollbar { display: none; }
+.jump .jump-title { display: none; margin: 0 0 .2rem; font-size: .75rem; font-weight: 700; letter-spacing: .08em;
+                    text-transform: uppercase; color: color-mix(in oklab, var(--brand) 72%, var(--ink)); }
+.jump a { flex: none; padding: .1rem 0; line-height: 1.15; color: var(--muted); text-decoration: none;
+          font-size: .9375rem; font-weight: 500; white-space: nowrap; }
+.jump a:hover, .jump a:focus-visible { color: var(--ink); }
+.jump a[aria-current] { color: var(--ink); font-weight: 700; box-shadow: inset 0 -.42em 0 var(--signal); }
+.panel.report { scroll-margin-top: 4.2rem; }        /* a jump lands under the bar, not behind it */
+@media (min-width: 75rem) {
+  #tab-suite:has(> .jump) { display: grid; grid-template-columns: minmax(0, 1fr) 11rem; }
+  #tab-suite:has(> .jump) > * { grid-column: 1; }
+  #tab-suite > .jump { grid-column: 2; grid-row: 1 / span 40; align-self: start; top: 1rem;
+                       flex-direction: column; align-items: flex-start; gap: .6rem; overflow: visible;
+                       margin: 1.4rem 1rem 1.4rem 0; padding: 1rem 1.1rem; border-radius: 14px; box-shadow: none; }
+  #tab-suite > .jump .jump-title { display: block; }
+  #tab-suite > .jump a { white-space: normal; }
+  .panel.report { scroll-margin-top: 1rem; }
+}
 /* Models: each connection a fold, its heading the summary */
 .model-card > summary h3.chart-head { display: inline; margin: 0; }
+/* Expand all / Collapse all: a text toggle with a chevron that turns, like a fold's own marker */
+button.fold-all { display: inline-flex; align-items: center; gap: .4rem; margin-left: auto; border: 0; padding: .2rem 0;
+                  background: none; color: var(--link); font: inherit; font-size: .9375rem; cursor: pointer; }
+button.fold-all:hover, button.fold-all:focus-visible { text-decoration: underline; text-underline-offset: .18em; }
+.fold-chev { width: .5rem; height: .5rem; border: solid currentColor; border-width: 0 2px 2px 0;
+             transform: rotate(-45deg); transition: transform .15s; }
+button.fold-all[aria-expanded="true"] .fold-chev { transform: rotate(45deg); margin-top: -.3rem; }
+@media (prefers-reduced-motion: reduce) { .fold-chev { transition: none; } }
 /* the Collections tab: the dialog's form, kept to a readable width */
 .panel.form-panel > :not(#c-list, table) { max-width: 46rem; }
 /* a list of things you open (collections), with its New button above */
@@ -595,7 +628,6 @@ dialog .pick { display: flex; gap: 1.5rem; }
   .toolbar input[type=search] { min-width: 0; width: 100%; }
   button.c-row { grid-template-columns: minmax(0, 1fr); }     /* the seats get the width; the whole row opens */
   .c-row .open-hint { display: none; }
-  .a-line { margin: 1rem 1rem 0; }
 }
 /* A narrow page (a phone, or a window with the sidebar beside it): name and
    state on one line, the words and actions under them. No gap between name

package/ui/bench.js

@@ -213,7 +213,7 @@
   var rendered = document.querySelector('.tab:not([data-lazy])');
   Promise.resolve().then(function () { if (rendered) setUp(rendered.id.slice(4)); });
   var wanted = store('bench.tab');            // else the one rendered: storage blocked or cleared, the cookie kept
-  show(wanted && $('tab-' + wanted) ? wanted : rendered ? rendered.id.slice(4) : 'glance');
+  show(wanted && $('tab-' + wanted) ? wanted : rendered ? rendered.id.slice(4) : 'suite');
 
   // ---- sortable tables -------------------------------------------------
   function key(cell) {
@@ -276,7 +276,88 @@
   document.addEventListener('input', function (ev) { if (ev.target.matches('input[data-filter]')) applyFilter(scopeOf(ev.target)); });
   document.addEventListener('change', function (ev) { if (ev.target.matches('select[data-facet]')) applyFilter(scopeOf(ev.target)); });
 
-  // ---- a link to another page (At a glance's line to Action needed) ----
+  // ---- Expand all / Collapse all: every fold of the page's groups (Models' connections).
+  // One labelled toggle, as accordions have it: Collapse all while every one is open,
+  // else Expand all. Kept right after any fold opens or shuts, by hand, by it or by a live swap.
+  var FOLDS = 'details.model-card, details[data-log]';      // the groups a page folds: Models' connections, Logs' logs
+  function foldButtons(tab) {
+    var folds = tab.querySelectorAll(FOLDS), open = 0;
+    folds.forEach(function (d) { if (d.open) open++; });
+    var all = open === folds.length;
+    tab.querySelectorAll('[data-fold-all]').forEach(function (b) {
+      b.dataset.foldAll = all ? 'close' : 'open';
+      b.setAttribute('aria-expanded', String(all));
+      b.lastChild.textContent = all ? 'Collapse all' : 'Expand all';
+    });
+  }
+  document.addEventListener('click', function (e) {
+    var b = e.target.closest('[data-fold-all]');
+    if (!b) return;
+    var tab = b.closest('.tab');
+    tab.querySelectorAll(FOLDS).forEach(function (d) { d.open = b.dataset.foldAll === 'open'; });
+    foldButtons(tab);
+  });
+  document.addEventListener('toggle', function (e) {            // toggle does not bubble: caught on its way down
+    if (e.target.matches && e.target.matches(FOLDS)) foldButtons(e.target.closest('.tab'));
+  }, true);
+
+  // ---- Logs: Export as text (every log in one .txt, to read) and CSV (one log, its
+  // own columns, for a spreadsheet). Both as the page shows them, the filter applied.
+  function stampNow() {                                     // local time, as the page shows
+    var now = new Date(), two = function (n) { return String(n).padStart(2, '0'); };
+    return now.getFullYear() + '-' + two(now.getMonth() + 1) + '-' + two(now.getDate()) + '-' + two(now.getHours()) + two(now.getMinutes());
+  }
+  // A log as the page shows it (the filter applied): its header and rows, as cells.
+  // A log of lines (Scheduler, Onboarding runs) is one column, Line.
+  function logRows(d) {
+    var cells = function (tr) {
+      return Array.prototype.map.call(tr.children, function (c) { return c.textContent.trim().replace(/\s+/g, ' '); });
+    };
+    var head = d.querySelector('thead tr'), rows = [head ? cells(head) : ['Line']];
+    d.querySelectorAll(head ? 'tbody tr' : '.log-line').forEach(function (r) {
+      if (!r.hidden) rows.push(head ? cells(r) : [r.textContent.trim()]);
+    });
+    return rows;
+  }
+  // Export…: the popup picks the format and the logs. Text: every log chosen in one
+  // .txt to read, tab-separated. CSV: a .csv per log (its own columns), through Share's
+  // saveFile(blob, name) and csv(rows) (RFC 4180, with a BOM).
+  on('click', 'logs-export', function () {
+    var q = ($('tab-logs').querySelector('input[data-filter]') || {}).value;
+    $('lx-note').textContent = q ? 'Only the lines that match the filter, "' + q + '", as the page shows them.'
+      : 'Every line the page shows: the last 200 of each log.';
+    $('lx-errors').textContent = '';
+    $('logs-export-dialog').showModal();
+  });
+  on('click', 'lx-cancel', function () { $('logs-export-dialog').close(); });
+  on('click', 'lx-go', function () {
+    var tab = $('tab-logs'), dlg = $('logs-export-dialog');
+    var format = dlg.querySelector('input[name=lx-format]:checked').value;
+    var want = Array.prototype.filter.call(dlg.querySelectorAll('input[name=lx-log]'), function (x) { return x.checked; })
+      .map(function (x) { return x.value; });
+    if (!want.length) { $('lx-errors').textContent = 'Choose at least one log.'; return; }
+    var logs = Array.prototype.filter.call(tab.querySelectorAll('details[data-log]'), function (d) {
+      return want.indexOf(d.dataset.log) !== -1;
+    });
+    var stamp = stampNow(), q = (tab.querySelector('input[data-filter]') || {}).value;
+    if (format === 'csv') {
+      logs.forEach(function (d) {
+        var slug = d.dataset.log.toLowerCase().replace(/[^a-z0-9]+/g, '-');
+        saveFile(new Blob([csv(logRows(d))], {type: 'text/csv;charset=utf-8'}), 'benchwarmers-' + slug + '-' + stamp + '.csv');
+      });
+    } else {
+      var out = ['Benchwarmers logs, exported ' + new Date().toString()];
+      if (q) out.push('Filtered by: ' + q);
+      logs.forEach(function (d) {
+        out.push('', '== ' + d.dataset.log + ' ==');
+        logRows(d).forEach(function (r) { out.push(r.join('\t')); });
+      });
+      saveFile(new Blob([out.join('\n') + '\n'], {type: 'text/plain;charset=utf-8'}), 'benchwarmers-logs-' + stamp + '.txt');
+    }
+    dlg.close();
+  });
+
+  // ---- a link to another page ----
   document.addEventListener('click', function (e) {
     var b = e.target.closest('[data-go]');
     if (b) { show(b.dataset.go); scrollTo(0, 0); }
@@ -767,8 +848,8 @@
   });
 
   // ---- after an action: the part it changed, not the whole page ----------
-  // Action needed's page, its count in the sidebar and At a glance's line
-  // about it, from the server (GET /api/needs, a few ms): a page reload
+  // Action needed's page and its count in the sidebar, from the server
+  // (GET /api/needs, a few ms): a page reload
   // renders every tab, seconds.
   var needsAsked = 0;
   function refreshNeeds(live) {
@@ -783,10 +864,6 @@
         count.textContent = n.count;
         tab.setAttribute('aria-label', tab.firstChild.textContent + ', ' + n.count + ' open');
       }
-      var line = document.querySelector('#tab-glance > .a-line');
-      if (line) line.remove();
-      // Not into a page not fetched yet: it says Loading… while empty, and comes with its line.
-      if (n.line && $('tab-glance').firstChild) $('tab-glance').insertAdjacentHTML('afterbegin', n.line);
     }).catch(function () { location.reload(); });          // could not ask: the old way
   }
   // A retired model's rows grey out and hide behind Show retired, as the
@@ -2125,7 +2202,10 @@
       var first = fact(g.name), caps = first.caps || {};
       var line = el('div', {class: 'm-row', role: 'group', 'aria-label': g.name, 'data-family': g.family, 'data-name': g.name});
       var name = el('div', {class: 'm-name'});
-      name.appendChild(el('a', {href: '#', class: 'model mono', 'data-alias': g.name, title: 'Open ' + g.name}, g.name));
+      // The same link as every model name elsewhere (a.model > .alias: the highlight on hover).
+      var link = el('a', {href: '#', class: 'model', 'data-alias': g.name, title: 'Open ' + g.name});
+      link.appendChild(el('span', {class: 'alias'}, g.name));
+      name.appendChild(link);
       line.appendChild(name);
       line.appendChild(el('span', {class: 'm-ctx'}, first.context ? tokens(first.context) : '—'));
       var can = el('span', {class: 'm-can'});
@@ -2683,6 +2763,9 @@
     // the pages share. It was the whole page, which made every chart wider than
     // its column beside the sidebar, scaled down, with smaller text.
     function width() {
+      // Benchmarks shown with its rail beside the reports: a report's own width, not the column's.
+      var report = document.querySelector('.tab.active .panel.report');
+      if (report && report.clientWidth) return Math.max(520, report.clientWidth - 50);
       var col = document.querySelector('.panes') || document.querySelector('main');
       return Math.max(520, (col.clientWidth || 900) - 50);
     }
@@ -2874,8 +2957,14 @@
           ]
         });
       };
-      put('chart-uptime-providers', function (w) { return uptime(U.providers, w); });
-      put('chart-uptime-models', function (w) { return uptime(U.models, w); });
+      // One view at a time, as the card's By says: by provider (the default) or by model.
+      var view = (document.querySelector('select[data-uptime-by]') || {}).value || 'providers';
+      ['providers', 'models'].forEach(function (k) {
+        var box = $('chart-uptime-' + k);
+        if (!box) return;
+        box.hidden = k !== view;
+        if (k === view) put('chart-uptime-' + k, function (w) { return uptime(U[k], w); });
+      });
     }
     if (part === 'uptime') return;
     // Decision models: each model's answers grouped by the probability it gave
@@ -3198,41 +3287,59 @@
       });
     });
 
-    // Vision, real work: each model's seats on image panels by how they were judged.
+    // Vision, real work: read as Real work is. The one question: when its seat was judged,
+    // did it help? The dot is the share judged useful (checked and helped, or useful with no
+    // claims to check), sized by how many were judged; the line where that share likely sits
+    // (95% Wilson), ranked by its low end. Beside it, in words: how often it answered at all,
+    // and the claims checked by hand. Seats never judged or never answered say nothing of
+    // its worth: they are in the words and the table, not the dot.
     put('chart-vision-work', function (w) {
-      var W = C.vision.work, kinds = C.vision.seatKinds, data = [], total = {};
+      var W = C.vision.work, kinds = C.vision.seatKinds, R = [], none = [];
       Object.keys(W).forEach(function (m) {
-        kinds.forEach(function (k) { if (W[m][k]) data.push({model: m, kind: k, count: W[m][k]}); total[m] = (total[m] || 0) + W[m][k]; });
-      });
-      if (!data.length) return el('p', {class: 'hint'}, 'No vision panels through the gateway yet.');
-      // The models doing the most image work first: their bars are the evidence. Under FEW
-      // seats a row was mostly "no answer" and said little: those are listed under the chart.
-      var all = Object.keys(total).sort(function (a, b) { return total[b] - total[a] || (a < b ? -1 : 1); });
-      var models = all.filter(function (m) { return total[m] >= FEW; });
-      var few = all.filter(function (m) { return total[m] < FEW; });
-      if (!models.length) { models = all; few = []; }
-      data = data.filter(function (d) { return models.indexOf(d.model) !== -1; });
-      var colour = {'checked, helped': good, 'checked, did not help': err, 'useful, no claims': sky,
-                    'no use, no claims': warn, 'not judged': line, 'no answer': muted};
-      // The claims checked by hand, beside the bar: the only right-or-wrong count real work has.
-      var claims = models.filter(function (m) { return W[m].real + W[m].false; }).map(function (m) {
-        return {model: m, text: W[m].real + ' of ' + (W[m].real + W[m].false) + ' claims right'};
+        var v = W[m], n = function (k) { return v[k] || 0; }, total = 0;
+        kinds.forEach(function (k) { total += n(k); });
+        if (!total) return;
+        var judged = n('checked, helped') + n('checked, did not help') + n('useful, no claims') + n('no use, no claims');
+        var useful = n('checked, helped') + n('useful, no claims');
+        var r = {model: m, total: total, judged: judged, useful: useful, answered: total - n('no answer'),
+                 real: v.real || 0, wrongC: v.false || 0};
+        if (!judged) { none.push(r); return; }
+        var b = wilson(useful, judged);
+        r.p = useful / judged; r.lo = b[0]; r.hi = b[1];
+        r.verdict = b[0] >= 0.5 ? 'helps' : b[1] < 0.5 ? 'does not help' : 'too few to say';
+        R.push(r);
       });
+      if (!R.length && !none.length) return el('p', {class: 'hint'}, 'No vision panels through the gateway yet.');
+      R.sort(function (a, b) { return b.lo - a.lo || b.judged - a.judged; });
       var narrowV = w < 560, box = el('div');
-      box.appendChild(Plot.plot({
-        width: w, height: rows(models.length), style: style, marginLeft: narrowV ? 150 : 260, marginRight: narrowV ? 20 : 150,
-        x: {label: 'Share of its seats on image panels →', tickFormat: '.0%', grid: true, domain: [0, 1]},
-        y: {domain: models, label: null, tickFormat: function (m) { return fit(m, narrowV ? 16 : 30) + '  (' + total[m] + ')'; }},
-        color: {legend: true, domain: kinds, range: kinds.map(function (k) { return colour[k]; })},
+      function words(r) {
+        return r.useful + '/' + r.judged + (narrowV ? '' : ' · ' + Math.round(r.answered / r.total * 100) + '% answered'
+          + (r.real + r.wrongC ? ' · claims ' + r.real + '/' + (r.real + r.wrongC) : ''));
+      }
+      if (R.length) box.appendChild(Plot.plot({
+        width: w, height: 24 * R.length + 90, style: style, marginLeft: narrowV ? 130 : 190,
+        marginRight: narrowV ? 50 : 260, marginTop: 34,
+        x: {domain: [0, 1], label: 'Judged useful, share of its judged seats →', tickFormat: '.0%', ticks: 5, grid: true},
+        y: {domain: R.map(function (r) { return r.model; }), label: null,
+            tickFormat: function (m) { return fit(m, narrowV ? 16 : 26); }},
+        r: {domain: [0, Math.max.apply(null, R.map(function (r) { return r.judged; }))], range: [3, 11]},
+        color: {domain: ['helps', 'too few to say', 'does not help'], range: [good, muted, err], legend: true},
         marks: [
-          Plot.barX(data, Plot.stackX({x: 'count', y: 'model', fill: 'kind', offset: 'normalize', order: kinds,
-            title: function (d) { return d.model + ': ' + d.kind + ', ' + d.count + ' of ' + total[d.model] + ' seats'; }})),
-          Plot.text(narrowV ? [] : claims.filter(function (c) { return models.indexOf(c.model) !== -1; }),
-            {x: 1, y: 'model', text: 'text', dx: 8, textAnchor: 'start', fill: ink})
+          Plot.ruleX([0.5], {stroke: line}),
+          Plot.ruleY(R, {y: 'model', x1: 'lo', x2: 'hi', stroke: 'verdict', strokeWidth: 2, strokeOpacity: 0.55}),
+          Plot.dot(R, {x: 'p', y: 'model', r: 'judged', fill: 'verdict', stroke: halo, strokeWidth: 1.5,
+            title: function (r) {
+              return r.model + '\njudged useful ' + r.useful + ' of ' + r.judged + ' judged seats ('
+                + Math.round(r.p * 100) + '%), likely between ' + Math.round(r.lo * 100) + '% and ' + Math.round(r.hi * 100)
+                + '%\nanswered ' + r.answered + ' of ' + r.total + ' seats on image panels'
+                + (r.real + r.wrongC ? '\nclaims checked by hand: ' + r.real + ' right of ' + (r.real + r.wrongC) : '');
+            }}),
+          Plot.text(R, {x: 1, y: 'model', dx: 10, textAnchor: 'start', fill: ink, fontSize: 11, text: words})
         ]
       }));
-      if (few.length) box.appendChild(el('p', {class: 'hint'}, 'Under ' + FEW + ' seats, so no bar: '
-        + few.map(function (m) { return m + ' (' + total[m] + ')'; }).join(', ') + '. The table has them all.'));
+      if (none.length) box.appendChild(el('p', {class: 'hint'}, 'No seat judged yet, so no dot: '
+        + none.map(function (r) { return r.model + ' (' + r.total + ' seats, answered ' + r.answered + ')'; }).join(', ')
+        + '. The table has every seat.'));
       return box;
     });
 
@@ -3271,10 +3378,50 @@
     });
 
   }
-  onLoad.suite = function () { drawCharts('suite'); };
+  onLoad.suite = function () { drawCharts('suite'); jumpSpy(); };
+  // Benchmarks' jump bar: the report in view is marked (aria-current) and its link kept in
+  // sight in a bar that scrolls sideways. Set up again after a live swap (setUp, onLoad).
+  var spy = null;
+  function jumpSpy() {
+    if (spy) spy.disconnect();
+    var nav = document.querySelector('#tab-suite .jump');
+    if (!nav || !window.IntersectionObserver) return;
+    var links = {}, seen = {}, current = null;
+    nav.querySelectorAll('a').forEach(function (a) { links[a.getAttribute('href').slice(1)] = a; });
+    // A report counts as in view while it crosses a thin band 15-20% down the window: just
+    // under where a jump lands it, so the one jumped to is the one marked.
+    spy = new IntersectionObserver(function (entries) {
+      entries.forEach(function (e) { seen[e.target.id] = e.isIntersecting; });
+      var id = Object.keys(links).filter(function (k) { return seen[k]; })[0];
+      if (!id || id === current) return;
+      if (current) links[current].removeAttribute('aria-current');
+      current = id;
+      links[id].setAttribute('aria-current', 'true');
+      var a = links[id];
+      if (a.offsetLeft < nav.scrollLeft || a.offsetLeft + a.offsetWidth > nav.scrollLeft + nav.clientWidth) {
+        nav.scrollLeft = a.offsetLeft - 16;
+      }
+    }, {rootMargin: '-15% 0px -80% 0px'});
+    document.querySelectorAll('#tab-suite .panel.report').forEach(function (p) { spy.observe(p); });
+  }
+  // A jump scrolls smoothly (unless motion is reduced) and moves focus to the report, so
+  // a keyboard user carries on from there; the address keeps the report's #, as a link would.
+  document.addEventListener('click', function (e) {
+    var a = e.target.closest('.jump a');
+    if (!a) return;
+    var target = $(a.getAttribute('href').slice(1));
+    if (!target) return;
+    e.preventDefault();
+    var still = window.matchMedia && matchMedia('(prefers-reduced-motion: reduce)').matches;
+    target.scrollIntoView({behavior: still ? 'auto' : 'smooth', block: 'start'});
+    target.setAttribute('tabindex', '-1');
+    target.focus({preventScroll: true});
+    history.replaceState(null, '', a.getAttribute('href'));
+  });
   // Uptime's or Trend's window: that window's tables (and Uptime's list) shown, its charts drawn again.
   document.addEventListener('change', function (ev) {
     var s = ev.target, part = s.matches('select[data-uptime-days]') ? 'uptime' : s.matches('select[data-trend-days]') ? 'trend' : '';
+    if (s.matches('select[data-uptime-by]')) { drawCharts('uptime'); return; }      // the other view, same window
     if (!part) return;
     s.closest('.panel').querySelectorAll('[data-' + part + '-window]').forEach(function (d) {
       d.hidden = d.getAttribute('data-' + part + '-window') !== s.value;