evals-lab 0.12.0 → 0.13.0

package/lab/metrics/builtin.mjs

@@ -139,9 +139,9 @@ function verdictOf(said        )
     can hold one. A reply that still cannot be read is asked for once more;
     a second, or a grader that failed, is the grader's error -- said as one,
     so it is not read as the reply failing. [ok] says whether a verdict read
-    is one this metric can use. */
+    is one this metric can use: true, or why not. */
 async function graded(ctx           , prompt        , schema                         ,
-                      ok                                               = () => true)                                        {
+                      ok                                                        = () => true)                                        {
   if (!ctx.ask) throw new Error("a model-graded metric needs a grader, and this run has none");
   let last = "";
   for (let attempt = 0; attempt < 2; attempt++) {
@@ -149,14 +149,21 @@ async function graded(ctx           , prompt        , schema
     try { said = await ctx.ask(prompt, schema); } catch (e) { throw new Error(`grader error: ${e instanceof Error ? e.message : String(e)}`); }
     try {
       const v = verdictOf(said);
-      if (ok(v)) return v;
-      last = `the grader said ${short(JSON.stringify(v))}`;
+      const usable = ok(v);
+      if (usable === true) return v;
+      last = typeof usable === "string" ? usable : `the grader said ${short(JSON.stringify(v))}`;
     } catch (e) { last = e instanceof Error ? e.message : String(e); }
   }
   throw new Error(`grader error: ${last}`);
 }
+/** Whether a pass-or-fail verdict's score is on the scale it was asked for.
+    One off it -- a 4 echoing a 1-to-5 rating -- is not clamped to a score the
+    grader never gave: it is read as unreadable, asked for again, then the
+    grader's error (#514). */
+const onScale = (v                              )                =>
+  typeof v.score !== "number" || (v.score >= 0 && v.score <= 1) || `the grader scored ${v.score}, off the 0 to 1 scale`;
 const gradedPass = (v                              , threshold        )                => {
-  const score = typeof v.score === "number" ? Math.max(0, Math.min(1, v.score)) : v.pass === true ? 1 : 0;
+  const score = typeof v.score === "number" ? v.score : v.pass === true ? 1 : 0;
   const pass = typeof v.pass === "boolean" ? v.pass && score >= threshold : score >= threshold;
   return { pass, score, reason: str(v.reason) || (pass ? "the grader passed it" : "the grader failed it") };
 };
@@ -339,6 +346,7 @@ const metrics                              = {
   // rules discarded is the finding.
   discarded: {
     family: RESULT,
+    readsResult: true,
     label: "Discarded",
     description: "Passes only when the job's rules discarded the reply.",
     options: [],
@@ -348,8 +356,26 @@ const metrics                              = {
       return ok(thrown, thrown ? "discarded" : input.error ? input.error : `kept ${input.terms.length} items`);
     },
   },
+  // A case that expects no answer at all: the connection answering an error
+  // -- a 400 for a file it can't read -- is the finding. A discard is
+  // Discarded's, at whichever stage it was.
+  refused: {
+    family: RESULT,
+    readsResult: true,
+    readsRefusal: true,
+    label: "Refused",
+    description: "Passes only when the target's call was refused: an error from the connection, not a reply.",
+    options: [],
+    defaults: () => ({}),
+    score: (input) => {
+      if (!input.error) return ok(false, "replied");
+      const refused = !/^(stage \d+: )?discarded: /.test(input.error);
+      return ok(refused, refused ? `refused: ${input.error}` : input.error);
+    },
+  },
   "item-count": {
     family: RESULT,
+    readsResult: true,
     label: "Item count",
     description: "Passes when the number of items is within the bounds.",
     options: [{ key: "min", label: "At least", type: "number" }, { key: "max", label: "At most", type: "number" }],
@@ -385,6 +411,7 @@ const metrics                              = {
   },
   latency: {
     family: RESULT,
+    readsResult: true,
     label: "Latency",
     description: "Passes when the reply came back within the time set.",
     options: [{ key: "max", label: "At most (ms)", type: "number" }],
@@ -472,7 +499,7 @@ const metrics                              = {
     validate: (m, at, bad) => needs(m, ["rubric"], at, bad),
     score: async (input, m, ctx) => gradedPass(await graded(ctx,
       `You are grading an output against a rubric.\n\n${taskOf(input, "")}Rubric:\n${str(m.rubric)}\n\nOutput:\n${input.text}\n\n${REPLY}`,
-      VERDICT_SCHEMA), num(m.threshold) ?? 0.5),
+      VERDICT_SCHEMA, onScale), num(m.threshold) ?? 0.5),
   },
   factuality: {
     family: GRADED,
@@ -488,7 +515,7 @@ const metrics                              = {
     score: async (input, m, ctx) => gradedPass(await graded(ctx,
       `You are checking an output for factual consistency with a reference. Differences in wording or detail are `
       + `fine; a contradiction, or a fact the reference does not support, is not.\n\n${taskOf(input, "")}Reference:\n`
-      + `${str(m.reference).trim() || recorded(input)}\n\nOutput:\n${input.text}\n\n${REPLY}`, VERDICT_SCHEMA), num(m.threshold) ?? 0.5),
+      + `${str(m.reference).trim() || recorded(input)}\n\nOutput:\n${input.text}\n\n${REPLY}`, VERDICT_SCHEMA, onScale), num(m.threshold) ?? 0.5),
   },
   "judge-vs-production": {
     family: GRADED,

package/lab/run-evals.js

@@ -479,20 +479,24 @@ function prepare(bin, file, c) {
 
 // ---- The type → handler registry ------------------------------------------
 // Design §4. What a file *is* to a run is decided here and only here, so the
-// queue and the page never learn about file types, and pdf and xlsx are
-// later entries that touch nothing but this map.
+// queue and the page never learn about file types, and xlsx is a later
+// entry that touches nothing but this map.
 //
 // A file with no mapper is passed through anyway -- decoded as UTF-8 with
 // replacement and joined at {text} like the text handler, and flagged in the
-// transcript. Stating the failure mode beats hiding it: a PDF's text layer
-// is not extracted, so its streams arrive as arbitrary bytes, and a run
-// that read them "well" would be a finding about nothing.
+// transcript. Stating the failure mode beats hiding it: a binary format's
+// bytes arrive as mojibake, and a run that read them "well" would be a
+// finding about nothing.
 const HANDLERS = new Map();
 [
   // Images: prepared to the profile's pixel budget by ImageMagick and sent
   // as the image content part. The formats are server.py's FILE_TYPES
   // image half, so nothing accepted for upload has no executor to run it.
   ["image", "jpg", "jpeg", "png", "heic", "heif", "dng", "tif", "tiff"],
+  // Documents: the file's bytes, unprocessed, handed to an API that reads
+  // them whole (core.DOCUMENT_TYPES); a connection that cannot is refused
+  // before the run, by validatePipeline.
+  ["document", ...Object.keys(core.DOCUMENT_TYPES).map(e => e.slice(1))],
   // Text-ish: decoded UTF-8, joining the stage at {text} -- the text-content
   // semantics the canvas already has (core.textPrompt).
   ["text", "txt", "md", "csv"],
@@ -511,8 +515,8 @@ const UNMAPPED = "the file has no handler, so it was passed through as raw "
 
 /**
  * One file of the Source made into what the job consumes: an image as the
- * dataUrl at [conn]'s budget, or text at {text}. A file with no mapper
- * still runs, and says so.
+ * dataUrl at [conn]'s budget, a document as its bytes, or text at {text}. A
+ * file with no mapper still runs, and says so.
  */
 function itemContent(bin, file, conn) {
   const kind = HANDLERS.get(extOf(file));
@@ -520,6 +524,7 @@ function itemContent(bin, file, conn) {
     const { dataUrl, dims } = prepare(bin, file, conn);
     return { kind, dataUrl, text: null, flags: [] };
   }
+  if (kind === "document") return { kind, dataUrl: documentOf(file), text: null, flags: [] };
   // TextDecoder with fatal unset replaces malformed bytes with U+FFFD --
   // what "decoded as UTF-8 with replacement" means, for a mapped-text file
   // and an unmapped one alike.
@@ -527,6 +532,13 @@ function itemContent(bin, file, conn) {
   return { kind: kind ?? "unmapped", dataUrl: null, text, flags: kind ? [] : [UNMAPPED] };
 }
 
+/** A document file as the attachment a stage sends: its bytes, its media
+    type and its name. */
+function documentOf(file) {
+  const name = path.basename(file);
+  return core.documentUrl(name, core.documentTypeOf(name).media, fs.readFileSync(file).toString("base64"));
+}
+
 // ---- Reaching a model ----------------------------------------------------
 
 /**
@@ -737,9 +749,10 @@ async function ask(to, c, prompt, dataUrl, schema = null) {
 // this invocation did, so resume and re-run item agree with a fresh run.
 
 // What a file *is* to a run, the same registry §4 gives the graded set: an
-// image is prepared to the profile's pixel budget; a text file is decoded
-// UTF-8 and joined at {text}; anything else is passed through as UTF-8 with
-// replacement and flagged, so the failure mode is stated rather than hidden.
+// image is prepared to the profile's pixel budget; a document is sent whole;
+// a text file is decoded UTF-8 and joined at {text}; anything else is passed
+// through as UTF-8 with replacement and flagged, so the failure mode is
+// stated rather than hidden.
 const RUN_IMAGE_TYPES = new Set([".jpg", ".jpeg", ".png", ".heic", ".heif", ".dng", ".tif", ".tiff"]);
 const RUN_TEXT_TYPES = new Set([".txt", ".md", ".csv"]);
 
@@ -916,6 +929,7 @@ async function runSnapshot(o) {
     : files.map(name => {
         const ext = path.extname(name).toLowerCase();
         if (RUN_IMAGE_TYPES.has(ext)) return { name, kind: "image" };
+        if (core.documentTypeOf(name)) return { name, kind: "document" };
         // A record (flows/record.ts): one call of a Power Automate step, what
         // an HTTP Request is rebuilt from.
         if (ext === ".json") return { name, kind: "record" };
@@ -981,6 +995,11 @@ async function runSnapshot(o) {
           break;
         }
       }
+      if (item.kind === "document") {
+        const file = path.join(o.source, item.name);
+        if (!fs.existsSync(file)) { failed = `the file ${item.name} is not in the source`; break; }
+        dataUrl = documentOf(file);
+      }
       let record = null;
       if (item.kind === "record") {
         const file = path.join(o.source, item.name);
@@ -1450,7 +1469,7 @@ async function main() {
       found: s.found, missed: s.missed, invented: s.invented, count: s.count,
       metrics: s.metrics, watch: s.watch,
       terms: res.terms || [], ms: res.ms,
-      // What a file type was to this run -- "image", "text", "unmapped".
+      // What a file type was to this run -- "image", "document", "text", "unmapped".
       type: item?.kind ?? null,
       // What the model actually said. #248 wants a failure written out as a
       // task an agent can act on, and "missed water" without the reply that

package/lab/server.py

@@ -50,6 +50,7 @@ from pathlib import Path
 # Ollama's own address on this machine, unless OLLAMA_URL names another: the
 # deployed lab's is set in the Dockerfile, and no default names a host of its.
 OLLAMA = os.environ.get("OLLAMA_URL", "http://127.0.0.1:11434")
+
 # The password in front of everything but /healthz (#463). Unset means open,
 # as a lab run on a workstation always has been; the deployed one sets it,
 # because it keeps setups and API keys and can spend GPU time.
@@ -145,6 +146,16 @@ def host_of(url: str) -> str:
     return (urllib.parse.urlsplit(url).hostname or "").lower()
 
 
+def ollama_shown():
+    """OLLAMA as the page may show it, where a connection with no address
+    goes: no user or password, should OLLAMA_URL carry one."""
+    try:
+        u = urllib.parse.urlsplit(OLLAMA)
+        return urllib.parse.urlunsplit(u._replace(netloc=u.netloc.rpartition("@")[2]))
+    except ValueError:
+        return ""
+
+
 # The chat endpoint each connection type's request reaches, off the /v1
 # base -- evals-core.ts's CONNECTION_TYPES, mirrored because the server is
 # stdlib-only and cannot load it. Ollama's native /api/chat sits at the root;
@@ -311,6 +322,8 @@ FILE_TYPES = {
     ".dng": "image/x-adobe-dng",
     ".tif": "image/tiff",
     ".tiff": "image/tiff",
+    # Handed to the model whole, as a document (evals-core.ts DOCUMENT_TYPES).
+    ".pdf": "application/pdf",
     ".txt": "text/plain; charset=utf-8",
     ".md": "text/markdown; charset=utf-8",
     ".csv": "text/csv; charset=utf-8",
@@ -1212,9 +1225,20 @@ def workspace_of(store) -> str:
 
 
 # The `connection_shares` sentinels (phase 4 fills and enforces the table;
-# phase 1 only creates it empty): every workspace, and future ones.
+# phase 1 only creates it empty): every workspace, future ones, and none.
+# No rows is All, so a set that loses its last workspace (a purge, the Share
+# link's Undo, a stale id) holds SHARE_NONE rather than nothing (#506).
 SHARE_ALL = "*all"
 SHARE_NEW = "*new"
+SHARE_NONE = "*none"
+
+def hold_nowhere(db, conns) -> None:
+    """SHARE_NONE for each (kind, id) in [conns] left with no share row, which
+    would otherwise read as shared with every workspace."""
+    db.executemany("INSERT INTO connection_shares (kind, id, workspace) SELECT ?, ?, ? "
+                   "WHERE NOT EXISTS (SELECT 1 FROM connection_shares WHERE kind = ? AND id = ?)",
+                   [(k, i, SHARE_NONE, k, i) for k, i in conns])
+
 
 # How the `docs` table scopes by key (docs/workspaces.md): workflows are
 # per-workspace, profiles/tokens global, and promptlab.versions is split --
@@ -1519,7 +1543,8 @@ class Store:
     # An AI Connection or an Account (kind "profile"/"account", id the
     # connection's -- the Target profile's it kept -- or the account's) is
     # usable in a workspace by its share set in `connection_shares`: a row for
-    # that workspace, or the SHARE_ALL sentinel.
+    # that workspace, or the SHARE_ALL sentinel. SHARE_NONE is a row that
+    # matches no workspace, so a set holding only it is usable nowhere.
     # No row at all is the migration and new-connection default -- shared with
     # every workspace, so nothing stops running the moment workspaces exist
     # (open question 2, resolved All); narrowing a connection is adding the
@@ -1540,7 +1565,7 @@ class Store:
         """Every connection's share set, for the Connections menus: keyed by
         "<kind>:<id>", the workspace values as stored (workspace ids and the
         sentinels). A connection with no row is absent, which the page reads as
-        shared with All. No key is anywhere in this (#258)."""
+        shared with All; one with SHARE_NONE is shared with no workspace. No key is anywhere in this (#258)."""
         out: dict = {}
         with self.lock, closing(sqlite3.connect(self.path)) as db:
             for kind, cid, ws in db.execute("SELECT kind, id, workspace FROM connection_shares"):
@@ -1548,15 +1573,20 @@ class Store:
         return out
 
     def set_shares(self, kind: str, cid: str, workspaces) -> None:
-        """Replace a connection's share set. SHARE_ALL means every workspace,
-        so it is kept alone; only the sentinels and live workspace ids are let
-        in, so a stale id cannot linger in the table."""
+        """Replace a connection's share set. SHARE_ALL means every workspace
+        and SHARE_NONE none, so each is kept alone; only the sentinels and live
+        workspace ids are let in, so a stale id cannot linger in the table. A
+        set named but left empty by that is SHARE_NONE, not the All an empty
+        set reads as; an empty set asked for is a plain delete, so a failed
+        lab import's rollback leaves no row for a connection it never wrote."""
         with self.lock, closing(sqlite3.connect(self.path)) as db, db:
             valid = {r[0] for r in db.execute("SELECT id FROM workspaces WHERE trash IS NULL")}
             chosen = [w for w in dict.fromkeys(workspaces or [])
-                      if w in (SHARE_ALL, SHARE_NEW) or w in valid]
+                      if w in (SHARE_ALL, SHARE_NEW, SHARE_NONE) or w in valid]
             if SHARE_ALL in chosen:
                 chosen = [SHARE_ALL]
+            elif SHARE_NONE in chosen or (workspaces and not chosen):
+                chosen = [SHARE_NONE]
             db.execute("DELETE FROM connection_shares WHERE kind = ? AND id = ?", (kind, cid))
             db.executemany("INSERT INTO connection_shares (kind, id, workspace) VALUES (?, ?, ?)",
                            [(kind, cid, w) for w in chosen])
@@ -1565,14 +1595,18 @@ class Store:
         """Add or remove one workspace from a connection's share set -- the Run
         bar's Share link and its Undo (docs/workspaces.md). The link is offered
         only where the connection is actually narrowed, so adding a row is what
-        grants the blocked workspace its use and Undo takes it back."""
+        grants the blocked workspace its use and Undo takes it back -- to
+        SHARE_NONE when it was the last, as a set shared nowhere was."""
         with self.lock, closing(sqlite3.connect(self.path)) as db, db:
             if on:
                 db.execute("INSERT OR IGNORE INTO connection_shares (kind, id, workspace) "
                            "VALUES (?, ?, ?)", (kind, cid, ws))
+                db.execute("DELETE FROM connection_shares WHERE kind = ? AND id = ? AND workspace = ?",
+                           (kind, cid, SHARE_NONE))
             else:
                 db.execute("DELETE FROM connection_shares WHERE kind = ? AND id = ? AND workspace = ?",
                            (kind, cid, ws))
+                hold_nowhere(db, [(kind, cid)])
 
     def workspace_name(self, ws) -> str:
         """A workspace's name, for the refusal sentence the Run bar shows."""
@@ -1702,8 +1736,11 @@ class Workspaces:
                      if isinstance(shares, list)
                      else list(db.execute("SELECT kind, id FROM connection_shares WHERE workspace = ?",
                                            (SHARE_NEW,))))
+            grant = [(k, i) for k, i in grant if k and i]
             db.executemany("INSERT OR IGNORE INTO connection_shares (kind, id, workspace) VALUES (?, ?, ?)",
-                           [(k, i, wid) for k, i in grant if k and i])
+                           [(k, i, wid) for k, i in grant])
+            db.executemany("DELETE FROM connection_shares WHERE kind = ? AND id = ? AND workspace = ?",
+                           [(k, i, SHARE_NONE) for k, i in grant])
         return wid, None
 
     def rename(self, wid, name) -> tuple:
@@ -1843,7 +1880,11 @@ class Workspaces:
             db.execute("DELETE FROM queue WHERE workspace = ?", (ws,))
             db.execute("DELETE FROM runs WHERE workspace = ?", (ws,))
             db.execute("DELETE FROM docs WHERE workspace = ?", (ws,))
+            # A connection shared only here is shared nowhere after, not with
+            # All, which is what no rows would read as (#506).
+            held = list(db.execute("SELECT kind, id FROM connection_shares WHERE workspace = ?", (ws,)))
             db.execute("DELETE FROM connection_shares WHERE workspace = ?", (ws,))
+            hold_nowhere(db, held)
             db.execute("DELETE FROM workspaces WHERE id = ?", (ws,))
         return sids, rids
 
@@ -4411,8 +4452,8 @@ class Datasets:
 
     def save(self, did, version, body):
         """The body, written at [version] -- the one it began from. Returns
-        ({version}, None), or (None, error); a stale version's error carries
-        the current row."""
+        ({version, versions}, None), or (None, error); a stale version's error
+        carries the current row."""
         if type(version) is not int:
             return None, (400, "a save names the version it began from")
         # A body of an earlier version -- from a page loaded before this one --
@@ -4431,7 +4472,9 @@ class Datasets:
             self._keep(db, r, text)
             db.execute("UPDATE datasets SET body = ?, version = ?, updated_at = ? WHERE id = ?",
                        (text, version + 1, self._now(), did))
-        return {"version": version + 1}, None
+            # The save may have cut a version: a link pins the newest (#547).
+            newest = self._head(db, did)["n"]
+        return {"version": version + 1, "versions": newest}, None
 
     def remove(self, did):
         """Into the trash, at once; Undo restores it. Returns (token, None)."""
@@ -4863,8 +4906,8 @@ class FilterSets:
 
     def save(self, fid, version, body):
         """The body, written at [version] -- the one it began from. Returns
-        ({version}, None), or (None, error); a stale version's error carries
-        the current row."""
+        ({version, versions}, None), or (None, error); a stale version's error
+        carries the current row."""
         if type(version) is not int:
             return None, (400, "a save names the version it began from")
         body = upgrade_filter_set_body(body)
@@ -4881,7 +4924,9 @@ class FilterSets:
             self._keep(db, r, text)
             db.execute("UPDATE filter_sets SET body = ?, version = ?, updated_at = ? WHERE id = ?",
                        (text, version + 1, self._now(), fid))
-        return {"version": version + 1}, None
+            # The save may have cut a version: a link pins the newest (#547).
+            newest = self._head(db, fid)["n"]
+        return {"version": version + 1, "versions": newest}, None
 
     def remove(self, fid):
         """Into the trash, at once; Undo restores it. Returns (token, None)."""
@@ -5141,6 +5186,43 @@ def demo_pack() -> bytes:
     return buf.getvalue()
 
 
+# The connection types a prompt cannot be asked of (evals-core.ts: their
+# `answers` lacks "prompt", or they answer from the item and are no model).
+NO_WORDS = {"echo", "recorded", "http"}
+
+
+def pack_connection(ws):
+    """The AI Connection a pack's target answers on where it names one this
+    lab lacks (#550): the lab's grader, or else the first connection that is
+    asked words, either shared with workspace [ws]; None where there is none,
+    so the target keeps its placeholder and the Run bar says to add one."""
+    body = ((STORE.all().get("promptlab.connections") or {}).get("body") or {}) if STORE is not None else {}
+    words = [c for c in body.get("list") or []
+             if isinstance(c, dict) and isinstance(c.get("id"), str) and c["id"]
+             and c.get("type") not in NO_WORDS and STORE.shared("profile", c["id"], ws)]
+    return next((c for c in words if c["id"] == body.get("grader")), words[0] if words else None)
+
+
+def repoint_targets(doc: dict, ws) -> None:
+    """A pack pipeline's targets that name a connection this lab does not have,
+    pointed at the one pack_connection picks, its profile dropped with it."""
+    targets = doc.get("targets")
+    if not isinstance(targets, list) or STORE is None:
+        return
+    have = {c.get("id") for c in ((STORE.all().get("promptlab.connections") or {}).get("body") or {}).get("list") or []
+            if isinstance(c, dict)}
+    pick = None
+    for t in targets:
+        ref = t.get("connection") if isinstance(t, dict) else None
+        if not isinstance(ref, dict) or ref.get("id") in have:
+            continue
+        pick = pick or pack_connection(ws)
+        if pick is None:
+            return
+        t["connection"] = {"id": pick["id"], "name": pick.get("name") or pick["id"]}
+        t["profile"] = None
+
+
 def pack_requires_problem(manifest: dict, plugins: set) -> str:
     """Why this lab cannot take a pack its manifest describes, or ""."""
     req = manifest.get("requires") or {}
@@ -5533,6 +5615,7 @@ class Packs:
                     sid = src_ids.get(content["ref"].get("id")) or src_ids.get(content["ref"].get("name"))
                     if sid:
                         content["ref"] = {"id": sid, "name": SOURCES.get(sid)["name"]}
+                repoint_targets(doc, ws)
                 new_work.append((key, doc))
             def place(body):
                 body.setdefault("list", [])
@@ -7518,10 +7601,12 @@ CONNECTION_TYPES = {
                   "presencePenalty", "dryMultiplier", "dryBase", "dryAllowedLength", "grammar",
                   "imagePosition", "cachePrompt", "thinking", "rejectLoops"),
 }
-# Types that answer from the item and send no request (Echo, #97): no
-# address to reach, no key to carry, so the destination rules have nothing to
-# say about them. A plugin's type joins with `"local": true`.
-LOCAL_CONNECTIONS = {"echo"}
+# Types that answer from the item and send no request (Echo, #97; Recorded,
+# #299): no address to reach, no key to carry, so the destination rules have
+# nothing to say about them. A target step stands in for each now, so a stored
+# row of one is never exported (lab_export). A plugin's type joins with
+# `"local": true`.
+LOCAL_CONNECTIONS = {"echo", "recorded"}
 # The built-in types, kept apart so a plugin's can be taken away again.
 BUILTIN_CONNECTION_TYPES = dict(CONNECTION_TYPES)
 BUILTIN_LOCAL = set(LOCAL_CONNECTIONS)
@@ -7967,6 +8052,11 @@ class NoRedirects(urllib.request.HTTPRedirectHandler):
 # -- a host nobody put on the list -- and the relay reported 200. An
 # allow-list that an environment variable can route around is not one.
 OPENER = urllib.request.build_opener(urllib.request.ProxyHandler({}), NoRedirects)
+# urllib's own "Python-urllib/3.x" is refused outright by many public APIs
+# behind Cloudflare (#512), so every outbound request names the lab instead --
+# honestly, never a browser. A request's own User-Agent, such as an API call
+# Source's Headers, wins: urllib adds this only where none is set.
+OPENER.addheaders = [("User-Agent", "evals-lab")]
 
 
 class Handler(BaseHTTPRequestHandler):
@@ -8158,7 +8248,8 @@ class Handler(BaseHTTPRequestHandler):
             # its own, and only in a lab with a store to keep it in.
             return self._json(200, {"targetCap": TARGET_CAP, "microsoft": microsoft_config(),
                                     "microsoftEditable": MICROSOFT is not None and not M365_CLIENT_ID,
-                                    "dynamicProviders": dynamic_providers_config()})
+                                    "dynamicProviders": dynamic_providers_config(),
+                                    "ollama": ollama_shown()})
         if path == "/api/state":
             if STORE is None:
                 return self._send(404, b"not found", "text/plain")
@@ -8970,7 +9061,11 @@ class Handler(BaseHTTPRequestHandler):
                         if err:
                             return self._json(err[0], {"error": err[1]})
                         content["live"] = True
-                        content["files"] = [name]
+                        # The one capture, looped as a snapshot's files are
+                        # (evals-core.ts expand): Run first cannot cut below
+                        # one item, and Loop repeats it for reply variance (#513).
+                        loops = content.get("loops")
+                        content["files"] = [name] * (loops if type(loops) is int and loops >= 1 else 1)
                         content["revs"] = {name: "live:" + hashlib.sha256(
                             produced[name].encode("utf-8")).hexdigest()}
                     else:
@@ -9702,8 +9797,10 @@ def lab_export(only=None) -> dict:
     the profiles over them (or over none) and the grader if it is one of them,
     each share set narrowed to that workspace, the one the file names, and
     *new kept where the connection has it. A connection shared only with
-    workspaces in the trash is usable nowhere, and an empty share set would
-    import as All, so it is left out, and out as the grader too."""
+    workspaces in the trash, or with none, is usable nowhere, and an empty share set would
+    import as All, so it is left out, and out as the grader too. So is a
+    connection of a local type (Echo): a target step stands in for it, and the
+    lab it lands in offers that step already."""
     body = (STORE.all().get("promptlab.connections") or {}).get("body") or {}
     profiles = (STORE.all().get("promptlab.connectionProfiles") or {}).get("body") or {}
     every = WORKSPACES.list()
@@ -9716,6 +9813,9 @@ def lab_export(only=None) -> dict:
     for c in body.get("list") or []:
         if not isinstance(c, dict) or (only and not STORE.shared("profile", str(c.get("id") or ""), spaces[0]["id"])):
             continue
+        if c.get("type") in LOCAL_CONNECTIONS:
+            nowhere.add(c.get("id"))
+            continue
         entry = {k: v for k, v in c.items() if k not in ("key", "shares")}
         held = rows.get(f"profile:{c.get('id')}")
         if held and SHARE_ALL not in held:
@@ -9827,6 +9927,11 @@ def lab_problems(doc) -> str:
         if not c["shares"]:
             return (f"connection {c['id']}'s shares is empty; leave shares out for All, "
                     "or name a workspace or *new")
+        # *none is the store's own mark for a set that lost its last workspace;
+        # an export leaves such a connection out, so a file never carries it.
+        if SHARE_NONE in c["shares"]:
+            return (f"connection {c['id']}'s shares names *none; a lab file holds no connection "
+                    "shared with no workspace -- leave it out, or name a workspace or *new")
         lost = [s for s in c["shares"] if s not in (SHARE_ALL, SHARE_NEW) and s not in known]
         if lost:
             return f"connection {c['id']} is shared with {', '.join(lost)}, which neither the file nor this lab has"
@@ -9893,9 +9998,7 @@ def lab_import(doc, keys) -> tuple:
             lines.append(f"grader set to {doc['grader']}")
         # Only an added connection takes the file's share set: one the lab had
         # keeps its own, as everything else it had does. Set before the write,
-        # so the connection is never there and shared with All in between; a
-        # failed write puts back what the rows were, since its ids never landed.
-        before = STORE.shares()
+        # so the connection is never there and shared with All in between.
         for c in added:
             if "shares" in c:
                 STORE.set_shares("profile", c["id"], [s if s in (SHARE_ALL, SHARE_NEW) else wid_of[s]
@@ -9912,9 +10015,15 @@ def lab_import(doc, keys) -> tuple:
         ok, _ = STORE.write(writes, STORE.default_ws)
         if ok:
             break
+        # A failed write takes back the rows of each id that still never
+        # landed, as no rows at all (an empty set is a delete, never *none).
+        # One another writer landed in between keeps the rows it has now:
+        # taking them would leave it shared with All.
+        landed = {x.get("id") for x in ((STORE.all().get("promptlab.connections") or {}).get("body")
+                                        or {}).get("list") or [] if isinstance(x, dict)}
         for c in added:
-            if "shares" in c:
-                STORE.set_shares("profile", c["id"], before.get(f"profile:{c['id']}") or [])
+            if "shares" in c and c["id"] not in landed:
+                STORE.set_shares("profile", c["id"], [])
     else:
         return None, (409, "the lab's connections kept changing while they were imported; import again")
     said += lines

package/lab/web/dist/assets/gallery-DpaqOg9Z.js

@@ -0,0 +1,5 @@
+import{$t as e,An as t,Bt as n,Cn as r,Ct as i,Dt as a,En as o,Ft as s,Ht as c,It as l,Jt as u,Kt as d,Nt as f,Pt as p,Qt as m,St as ee,Tn as te,Tt as h,Ut as ne,Vt as g,Xt as _,Yt as v,_t as re,an as ie,at as ae,bn as oe,bt as se,cr as ce,fn as y,gn as le,gt as ue,kn as de,kt as b,l as fe,ln as pe,lt as x,mn as S,n as me,nn as C,on as he,pn as ge,qt as _e,s as w,sr as ve,tn as ye,vn as be,vt as T,wn as xe,wt as Se,xt as Ce,yn as we}from"./tokens-CrjDMCbz.js";var E=ce(),Te=ve(),D=[{id:`text`,label:`Plain text`,settings:[]},{id:`list`,label:`List`,settings:[{key:`parse`,label:`Parse as`,type:`select`,choices:[{value:`csv`,label:`CSV`},{value:`lines`,label:`One a line`},{value:`json`,label:`JSON array`}]}]}],O=[{id:`case`,label:`Case`,accepts:[`list`],options:[{key:`mode`,label:`Mode`,type:`select`,choices:[{value:`LOWER`,label:`lower`},{value:`SENTENCE`,label:`Sentence`},{value:`TITLE`,label:`Title`},{value:`UPPER`,label:`UPPER`},{value:`NONE`,label:`As replied`}]}],defaults:()=>({type:`case`,mode:`LOWER`})},{id:`join`,label:`Join`,accepts:[`list`],options:[{key:`with`,label:`Separator`,type:`text`}],defaults:()=>({type:`join`,with:`, `})}],Ee=[{name:`language`,type:`value`,value:`English UK`},{name:`geo-location`,type:`value`,value:`Hamilton, Waikato, New Zealand`},{name:`geo`,type:`block`,enabled:!1}],De=[{path:`group/open`,title:`Group, open`},{path:`group/closed`,title:`Group, closed with a note`},{path:`group/no-menu`,title:`Group without actions`},{path:`group/sub`,title:`Sub-group, level 3`},{path:`group/depth`,title:`Group, three levels deep`},{path:`group/renaming`,title:`Group, renaming`},{path:`menu/disabled`,title:`Menu with a disabled item`},{path:`row/button`,title:`Row actions, one action a button`},{path:`row/menu`,title:`Row actions, two or more in the menu`},{path:`actionbar/states`,title:`Action bar, docked under a header, in its four states`},{path:`tabs/main`,title:`Tab strip, the lab's tabs`},{path:`tabs/section`,title:`Tab strip, a section strip under the lab's`},{path:`header/pipeline`,title:`Context header, a Pipeline`},{path:`header/dataset`,title:`Context header, a Dataset`},{path:`header/source`,title:`Context header, a Source`},{path:`header/switcher`,title:`Switcher, recent and more`},{path:`header/switcher-detail`,title:`Switcher, with a detail`},{path:`picker/fields`,title:`Picker, a profile and a Source`},{path:`multi/closed`,title:`Multi-select, two of three ticked`},{path:`multi/open`,title:`Multi-select, its list open`},{path:`chiplist/words`,title:`Chip-list, words with the input to add more`},{path:`chiplist/empty`,title:`Chip-list, empty with its placeholder`},{path:`workspace/menu`,title:`Workspace trigger and its three-part menu`},{path:`copymove/choose`,title:`Copy or move, a destination chosen`},{path:`copymove/collision`,title:`Copy or move, the name is already taken`},{path:`copymove/confirm`,title:`Copy or move, a Move that breaks a dependency`},{path:`copymove/confirm-pipeline`,title:`Copy or move, a pipeline Move that leaves its data behind`},{path:`share/menu`,title:`Share menu, All / some / New workspaces`},{path:`header/long`,title:`Context header, a long name`},{path:`menu/busy`,title:`Menu with a busy item`},{path:`loading/group`,title:`Loading, inside a group`},{path:`loading/error`,title:`Loading failed, with Retry`},{path:`loading/busy-button`,title:`Busy buttons`},{path:`repeater/empty`,title:`Repeater, empty`},{path:`repeater/min`,title:`Repeater at min 1`},{path:`repeater/between`,title:`Repeater, two of four`},{path:`repeater/max`,title:`Repeater at max`},{path:`repeater/locked`,title:`Repeater, a locked row`},{path:`props/controls`,title:`Property table, every control`},{path:`props/required`,title:`Property table, required only`},{path:`props/long`,title:`Property table, long values`},{path:`props/cards`,title:`Property table, a kind chosen by card`},{path:`props/widths`,title:`Property table, a width per value`},{path:`props/affix`,title:`Property table, a fixed suffix and a field's error`},{path:`props/secret`,title:`Property table, a key held and not held`},{path:`props/code`,title:`Property table, an endpoint and a payload`},{path:`table/sortable`,title:`Table, sortable headers and a row menu`},{path:`table/paged`,title:`Table, paged ten at a time`},{path:`table/select`,title:`DataTable with selection`},{path:`table/select-on`,title:`Table, two rows selected`},{path:`table/filter`,title:`Table with a filter bar`},{path:`table/filter-on`,title:`Table, filtered, with a scope`},{path:`edit/form`,title:`Edit modal, a profile`},{path:`edit/plain`,title:`Edit modal, one section and no menu`},{path:`edit/long`,title:`Edit modal, a form that scrolls`},{path:`edit/status`,title:`Edit modal, a status that is a refusal`},{path:`view/tokens`,title:`Token mappings`},{path:`view/output`,title:`Output with Case`},{path:`view/output-text`,title:`Output, plain text`},{path:`view/output-two`,title:`Output with two modifiers`},{path:`view/output-twice`,title:`Output with the same modifier twice`},{path:`view/output-none`,title:`Responses, read as text`},{path:`view/output-flow`,title:`Responses with Read as`},{path:`view/field-chip`,title:`Field chips in a target's words, with a changed word`},{path:`view/field-edit`,title:`The Edited target's editable prompt, with Insert field and a highlighted diff`},{path:`view/insert-field`,title:`The Edited target's Insert field menu, open`},{path:`wizard/pill`,title:`Wizard pill, two of four steps done`},{path:`wizard/pill-open`,title:`Wizard pill, its checklist open on the current step`},{path:`switch/off`,title:`Switch, off`},{path:`switch/on`,title:`Switch, on`},{path:`switch/disabled`,title:`Switch, disabled`}],k=e(),A=()=>{},j=(...e)=>e.map(e=>({label:e,icon:e===`Remove`?m.remove:e===`Clone`?m.clone:m.rename,onSelect:A,destructive:e===`Remove`}));function M({start:e,min:t,max:n,locked:r}){let[i,a]=(0,E.useState)(e);return(0,k.jsx)(re,{noun:`scenario`,items:i,...t===void 0?{}:{min:t},...n===void 0?{}:{max:n},...r?{locked:r}:{},onAdd:()=>a([...i,`Target ${String.fromCharCode(65+i.length)}`]),onRemove:e=>a(i.filter((t,n)=>n!==e)),onClone:e=>a([...i.slice(0,e+1),`${i[e]} copy`,...i.slice(e+1)]),onRename:A,empty:`No targets yet.`,renderItem:(e,t,n)=>(0,k.jsx)(d,{id:`scenario-${t}`,title:e,level:3,actions:n,children:(0,k.jsx)(`p`,{className:`note`,children:`A connection, a profile and a prompt for each job.`})},`${e}-${t}`)})}function Oe(){let[e,t]=(0,E.useState)(Ee),n=(n,r)=>t(e.map(e=>e.name===n?{...e,...r}:e));return(0,k.jsx)(fe,{mappings:e,onValue:(e,t)=>n(e,{value:t}),onEnabled:(e,t)=>n(e,{enabled:t}),onEdit:A,onRemove:n=>t(e.filter(e=>e.name!==n)),onAdd:()=>t([...e,{name:`token${e.length+1}`,type:`value`,value:``}])})}function N({start:e,format:t=!0,readAs:n}){let[r,i]=(0,E.useState)(e),[a,o]=(0,E.useState)(t),[s,c]=(0,E.useState)(n);return(0,k.jsx)(ae,{id:`output`,out:r,format:a,readAs:s,outputKinds:D,modifiers:O,onFormat:e=>{o(e!=null);let t=e??`text`;i({kind:t,modifiers:r.modifiers.filter(e=>O.find(t=>t.id===e.type)?.accepts.includes(t))})},onReadAs:c,onModifier:(e,t)=>i({...r,modifiers:r.modifiers.map((n,r)=>r===e?t:n)}),onAddModifier:e=>i({...r,modifiers:[...r.modifiers,e]}),onRemoveModifier:e=>i({...r,modifiers:r.modifiers.filter((t,n)=>n!==e)})})}function ke(){let[e,t]=(0,E.useState)(`Job 1`),[n,r]=(0,E.useState)(!0);return(0,k.jsx)(d,{id:`renaming`,title:e,actions:[{label:`Rename…`,icon:m.rename,onSelect:()=>r(!0)}],...n?{rename:{value:e,onCommit:n=>{t(n||e),r(!1)},onCancel:()=>r(!1)}}:{},children:(0,k.jsx)(`p`,{className:`note`,children:`Enter or leaving keeps the name; Escape cancels.`})})}function Ae(){let[e,t]=(0,E.useState)(`system`);return(0,k.jsx)(`div`,{className:`gallery-row`,style:{justifyContent:`flex-end`},children:(0,k.jsx)(i,{current:{id:`w1`,slug:`work`,name:`Work`},workspaces:[{id:`w1`,slug:`work`,name:`Work`},{id:`w2`,slug:`side-project`,name:`Side project`},{id:`w3`,slug:`client-demos`,name:`Client demos`}],onSwitch:A,onNew:A,onManage:A,theme:e,onTheme:t})})}var je=[{id:`cur`,slug:`work`,name:`Work`,archived:!1,isDefault:!0,lastUsed:`2026-10-08T09:30:00Z`,createdAt:null,pipelines:4,runs:18},{id:`side`,slug:`side-project`,name:`Side project`,archived:!1,isDefault:!1,lastUsed:`2026-10-07T09:30:00Z`,createdAt:null,pipelines:2,runs:6},{id:`demos`,slug:`client-demos`,name:`Client demos`,archived:!1,isDefault:!1,lastUsed:`2026-10-06T09:30:00Z`,createdAt:null,pipelines:1,runs:0}],Me={id:`cur`,slug:`work`,name:`Work`},Ne={undo:null,offer:A,undoNow:A,dismiss:A},Pe=e=>({copyMovePreflight:async()=>({nameTaken:e,suggestedName:`Daily regressions (copy)`}),copyAsset:async()=>({id:`copy`}),moveAsset:async()=>({id:`same`}),forWorkspace:()=>({})});function P({nameTaken:e=!1,confirm:t=!1,subject:n,dependents:r=[]}){return(0,k.jsx)(ue,{api:Pe(e),subject:n??{kind:`filter-set`,id:`f1`,name:`Daily regressions`},workspaces:je,current:Me,dependents:r,undo:Ne,onClose:A,defaultDest:`side`,...t?{defaultConfirm:!0}:{}})}function Fe(){let e=[{id:`w1`,name:`Work`},{id:`w2`,name:`Side project`},{id:`w3`,name:`Client demos`}];return(0,k.jsx)(`div`,{style:{display:`grid`,gap:`var(--s3)`,maxWidth:`20rem`},children:[{label:`All workspaces`,shares:void 0},{label:`Some workspaces`,shares:[`w2`]},{label:`New workspaces`,shares:[se]},{label:`No workspaces`,shares:[Ce]}].map(t=>(0,k.jsx)(ee,{label:t.label,workspaces:e,shares:t.shares,onChange:A},t.label))})}function F({sections:e,menu:t,status:n,statusBad:r,outcome:i}){return(0,k.jsx)(f,{title:`Lab Ollama`,...n?{status:n}:{},...r?{statusBad:!0}:{},...t?{menu:t}:{},sections:e,actions:[{label:`Test connection`,icon:ge,onSelect:A,outcome:i?{label:i,onOpen:A}:void 0}],onClose:A})}var I=()=>[{key:`name`,label:`Name`,width:`full`,required:!0,control:{type:`text`,value:`Lab Ollama`,onChange:A}},{key:`slug`,label:`Slug`,width:`full`,required:!0,control:{type:`text`,value:`lab-ollama`,onChange:A}},{key:`type`,label:`Type`,required:!0,control:{type:`select`,value:`ollama`,onChange:A,options:[{value:`ollama`,label:`Ollama (Local)`},{value:`anthropic`,label:`Anthropic`}]}},{key:`url`,label:`Address`,width:`full`,control:{type:`text`,value:``,placeholder:`http://127.0.0.1:11434`,onChange:A}},{key:`key`,label:`Key`,control:{type:`secret`,value:``,onChange:A}},{key:`model`,label:`Model`,width:`full`,control:{type:`text`,value:`minicpm-v4.6:latest`,onChange:A}}],L=()=>[{key:`seed`,label:`Seed`,control:{type:`text`,value:``,placeholder:`1234`,onChange:A}},{key:`nPredict`,label:`Reply tokens`,width:`quarter`,control:{type:`text`,value:``,onChange:A}},{key:`temperature`,label:`Temperature`,control:{type:`text`,value:`0.3`,onChange:A}},{key:`repeatPenalty`,label:`Repeat penalty`,width:`quarter`,control:{type:`text`,value:``,onChange:A}}],R=()=>[{key:`endpoint`,label:`Endpoint`,required:!0,control:{type:`code`,text:`http://127.0.0.1:11434/api/chat`}},{key:`prompt`,label:`Prompt`,required:!0,control:{type:`static`,text:`(prompt)`}},{key:`model`,label:`model`,control:{type:`code`,text:`minicpm-v4.6:latest`}},{key:`options`,label:`options`,control:{type:`code`,text:`{"seed":1234,"temperature":0.3}`}}],z={key:`kind`,label:`Response Format`,required:!0,control:{type:`select`,value:`list`,onChange:A,options:D.map(e=>({value:e.id,label:e.label}))}};function B({actions:e}){return(0,k.jsx)(d,{id:`g`,title:`Actions`,children:(0,k.jsx)(b,{label:`Actions`,rows:[{name:`Ask_Claude`}],rowKey:e=>e.name,columns:[{key:`name`,header:`Name`,role:`title`,cell:e=>e.name},{key:`menu`,header:``,role:`menu`,sort:null,cell:()=>(0,k.jsx)(_e,{label:`Ask_Claude actions`,actions:e})}]})})}var Ie=[[`home`,`Home`,we],[`runs`,`Runs`,te],[`history`,`History`,oe],[`library`,`Library`,be],[`setup`,`Setup`,pe]],Le=[[`datasets`,`Datasets`,t],[`prompts`,`Prompts`,le],[`sources`,`Sources`,xe]],V=e=>(0,k.jsx)(g,{label:`Lab`,current:e,items:Ie.map(([e,t,n])=>({key:e,label:t,icon:n,href:`#/${e}`}))}),H=e=>(0,k.jsx)(g,{label:`Library`,level:`section`,current:e,items:Le.map(([e,t,n])=>({key:e,label:t,icon:n,href:`#/library/${e}`}))}),U=({section:e,children:t})=>(0,k.jsxs)(`div`,{children:[V(e?`library`:`runs`),e&&H(e),t]}),W=(...e)=>e.map(e=>({label:e,icon:e===`Delete`?m.remove:e===`Clone`?m.clone:m.rename,onSelect:A,destructive:e===`Delete`})),Re={"header/pipeline":()=>(0,k.jsx)(U,{children:(0,k.jsx)(n,{kind:`Pipeline`,name:`Pipeline 1`,onPick:A,meta:`2 jobs, 3 targets`,primary:{label:`New pipeline`,icon:y,onSelect:A},actions:W(`Rename…`,`Clone`,`Export YAML`,`Import YAML…`,`Restore…`,`Delete`)})}),"header/dataset":()=>(0,k.jsx)(U,{section:`datasets`,children:(0,k.jsx)(n,{kind:`Dataset`,name:`Demo 1`,onPick:A,meta:`62 cases, 19 graded`,primary:{label:`New dataset`,icon:y,onSelect:A},actions:W(`Rename…`,`Clone`,`Import JSON…`,`Export JSON`,`Restore…`,`Delete`)})}),"header/switcher":()=>(0,k.jsx)(h,{kind:`Pipeline`,value:`p5`,onPick:A,onClose:A,recent:[`p7`,`p6`,`p5`,`p4`,`p3`],options:Array.from({length:13},(e,t)=>({id:`p${t}`,label:[`Demo 1`,`Demo 2`,`Receipts`,`Menus`,`Street signs`,`Labels`,`Plates`,`Posters`,`Packaging`,`Whiteboards`,`Book spines`,`Shop fronts`,`Timetables`][t],used:t<8?{at:`2026-10-0${t+1}T09:30:00Z`,label:x(`2026-10-0${t+1}T09:30:00Z`)}:null}))}),"header/switcher-detail":()=>(0,k.jsx)(h,{kind:`Connection`,value:`c1`,onPick:A,onClose:A,detailLabel:`Model`,recent:[`c3`,`c1`],options:[[`Lab Ollama gemma3 4b`,`gemma3:4b`],[`Lab Ollama minicpm`,`minicpm-v`],[`Anthropic`,`claude-sonnet-4-5`],[`Dell llama`,`qwen2.5vl:7b`],[`OpenAI`,`gpt-5`],[`Docker llama`,`qwen2.5vl:7b`]].map(([e,t],n)=>({id:`c${n}`,label:e,detail:t,used:n<4?{at:`2026-10-1${n}T09:30:00Z`,label:x(`2026-10-1${n}T09:30:00Z`)}:null}))}),"picker/fields":()=>(0,k.jsx)(d,{id:`g`,title:`Target A`,children:(0,k.jsx)(p,{label:`Target A setup`,rows:[{key:`profile`,label:`Profile`,control:{type:`picker`,kind:`Profile`,value:`p2`,usedLabel:`Last changed`,onChange:A,recent:[`p5`,`p4`,`p3`,`p2`,`p1`],options:[{id:`#echo`,label:`Echo`,fixed:!0},...[`Lab Ollama`,`Anthropic`,`Cloud vision`,`Local llama`,`OpenRouter`,`Bench`,`Vision`].map((e,t)=>({id:`p${t}`,label:e,used:t<6?{at:`2026-10-0${t+1}T09:30:00Z`,label:x(`2026-10-0${t+1}T09:30:00Z`)}:null}))]}},{key:`source`,label:`Source`,control:{type:`picker`,kind:`Source`,value:`gone`,usedLabel:`Last changed`,onChange:A,options:[{id:``,label:`Any`,fixed:!0},{id:`gone`,label:`Shoot 3 (missing)`,missing:!0},{id:`s1`,label:`Receipts`}]}}]})}),"header/source":()=>(0,k.jsx)(U,{section:`sources`,children:(0,k.jsx)(n,{kind:`Source`,name:`Demo 1 images`,onPick:A,meta:`File Library. 63 files, 48 MB`,primary:{label:`Upload…`,icon:C,onSelect:A},actions:W(`Rename…`,`Delete`)})}),"workspace/menu":()=>(0,k.jsx)(Ae,{}),"copymove/choose":()=>(0,k.jsx)(P,{}),"copymove/collision":()=>(0,k.jsx)(P,{nameTaken:!0}),"copymove/confirm":()=>(0,k.jsx)(P,{confirm:!0,dependents:[{kind:`pipeline`,id:`w1`,name:`Nightly regression`,href:`#/runs`}]}),"copymove/confirm-pipeline":()=>(0,k.jsx)(P,{confirm:!0,subject:{kind:`pipeline`,id:`w1`,name:`Nightly regression`},dependents:[{kind:`eval-group`,id:`g1`,name:`Receipts, 62 cases`,href:`#/library/datasets/g1`},{kind:`source`,id:`s1`,name:`Receipt images`,href:`#/library/sources`}]}),"share/menu":()=>(0,k.jsx)(Fe,{}),"header/long":()=>(0,k.jsx)(U,{children:(0,k.jsx)(n,{kind:`Pipeline`,name:`Receipts and handwritten menus, second pass with the tighter prompt and a Contains metric`,onPick:A,meta:`12 jobs, 8 targets`,primary:{label:`New pipeline`,icon:y,onSelect:A},actions:W(`Rename…`,`Clone`,`Delete`)})}),"actionbar/states":()=>(0,k.jsxs)(`div`,{children:[V(`runs`),(0,k.jsx)(n,{kind:`Pipeline`,name:`Demo 1`,onPick:A,meta:`1 job, 2 targets`,docked:!0,primary:{label:`New pipeline`,icon:y,onSelect:A},actions:W(`Rename…`,`Delete`)}),(0,k.jsx)(T,{primary:{label:`Run`,icon:S,onSelect:A},status:(0,k.jsx)(`b`,{children:`Ready to run`})}),(0,k.jsxs)(`div`,{style:{display:`grid`,gap:`var(--s4)`,marginTop:`var(--s5)`},children:[(0,k.jsx)(T,{primary:{label:`Run`,icon:S,onSelect:A,disabled:!0},status:(0,k.jsxs)(k.Fragment,{children:[`Connection "Vision" not found. `,(0,k.jsxs)(`a`,{href:`#/setup`,style:{whiteSpace:`nowrap`},children:[`Open`,`\xA0`,`Setup`]})]})}),(0,k.jsx)(T,{primary:{label:`Run`,icon:S,onSelect:A,disabled:!0},status:(0,k.jsxs)(k.Fragment,{children:[(0,k.jsx)(`b`,{children:`Running`}),` 12 of 42`]}),secondary:{label:`Stop`,icon:he,onSelect:A},progress:28,progressLabel:`Run progress`}),(0,k.jsx)(T,{primary:{label:`Run`,icon:S,onSelect:A},status:(0,k.jsxs)(k.Fragment,{children:[(0,k.jsx)(`b`,{children:`Done`}),` 42 in 2m`]})})]})]}),"tabs/main":()=>V(`runs`),"tabs/section":()=>(0,k.jsxs)(`div`,{children:[V(`library`),H(`datasets`)]}),"group/open":()=>(0,k.jsx)(d,{id:`g`,title:`Evals`,actions:j(`Change…`,`Remove`),children:(0,k.jsx)(`p`,{className:`note`,children:`The graded set.`})}),"group/closed":()=>(0,k.jsx)(d,{id:`g`,title:`Content`,open:!1,note:`Sample library, 63 files`,actions:j(`Change…`,`Remove`)}),"group/no-menu":()=>(0,k.jsx)(d,{id:`g`,title:`Mapping`,children:(0,k.jsx)(`p`,{className:`note`,children:`Nothing to do to it.`})}),"group/sub":()=>(0,k.jsx)(d,{id:`g`,title:`Job 1`,actions:j(`Rename…`,`Remove`),children:(0,k.jsx)(d,{id:`g-sub`,title:`Scenarios`,level:3,actions:j(`Clear`),children:(0,k.jsx)(`p`,{className:`note`,children:`Inside a pane.`})})}),"group/depth":()=>(0,k.jsxs)(d,{id:`g`,title:`Job 1`,actions:j(`Rename…`,`Remove`),children:[(0,k.jsxs)(d,{id:`g-scenarios`,title:`Scenarios`,level:3,children:[(0,k.jsx)(d,{id:`g-plain`,title:`Plain`,level:3,actions:j(`Rename…`,`Remove`),children:(0,k.jsx)(`p`,{className:`note`,children:`A pane, a section of it, and a thing with a name inside that section.`})}),(0,k.jsx)(d,{id:`g-strict`,title:`Strict`,level:3,actions:j(`Rename…`,`Remove`)})]}),(0,k.jsx)(d,{id:`g-output`,title:`Output`,level:3,children:(0,k.jsx)(`p`,{className:`note`,children:`The next section of the same pane.`})})]}),"group/renaming":()=>(0,k.jsx)(ke,{}),"menu/disabled":()=>(0,k.jsx)(d,{id:`g`,title:`Target D`,actions:[{label:`Rename…`,icon:m.rename,onSelect:A},{label:`Clone`,icon:m.clone,onSelect:A,disabled:!0},{label:`Remove`,icon:m.remove,onSelect:A,destructive:!0}]}),"row/button":()=>(0,k.jsx)(B,{actions:[{label:`New pipeline`,icon:S,onSelect:A}]}),"row/menu":()=>(0,k.jsx)(B,{actions:[{label:`View…`,icon:o,onSelect:A},{label:`Download`,icon:de,onSelect:A},{label:`Remove`,icon:ie,onSelect:A,destructive:!0}]}),"menu/busy":()=>(0,k.jsx)(d,{id:`g`,title:`History`,actions:[{label:`Export JSON`,icon:m.export,onSelect:A,busy:!0},{label:`Clear`,icon:m.clear,onSelect:A,destructive:!0}]}),"loading/group":()=>(0,k.jsx)(d,{id:`g`,title:`History`,actions:j(`Rename…`),children:(0,k.jsx)(v,{label:`Runs`,delay:0})}),"loading/error":()=>(0,k.jsx)(d,{id:`g`,title:`History`,children:(0,k.jsx)(v,{label:`Runs`,error:`Runs not loaded: server unreachable.`,onRetry:A})}),"loading/busy-button":()=>(0,k.jsx)(d,{id:`g`,title:`Run`,children:(0,k.jsxs)(`div`,{className:`gallery-row`,children:[(0,k.jsxs)(`button`,{type:`button`,className:`run`,disabled:!0,"aria-busy":!0,children:[(0,k.jsx)(_,{size:16}),`Run`]}),(0,k.jsxs)(`button`,{type:`button`,className:`ghost`,disabled:!0,"aria-busy":!0,children:[(0,k.jsx)(_,{}),`Older`]})]})}),"repeater/empty":()=>(0,k.jsx)(M,{start:[]}),"repeater/min":()=>(0,k.jsx)(M,{start:[`Target A`],min:1}),"repeater/between":()=>(0,k.jsx)(M,{start:[`Target A`,`Target B`],min:1,max:4}),"repeater/max":()=>(0,k.jsx)(M,{start:[`Target A`,`Target B`,`Target C`,`Target D`],min:1,max:4}),"repeater/locked":()=>(0,k.jsx)(M,{start:[`Reply`,`Target A`],locked:(e,t)=>t===0}),"props/controls":()=>(0,k.jsx)(p,{label:`Controls`,onRename:A,onRemove:A,onAdd:A,addLabel:`Add row`,rows:[z,{key:`language`,label:`{language}`,code:!0,control:{type:`text`,value:`English UK`,onChange:A}},{key:`geo`,label:`{geo}…{/geo}`,code:!0,control:{type:`static`,text:`Block`},enabled:{checked:!0,onChange:A}},{key:`image`,label:`Image`,control:{type:`checkbox`,checked:!0,label:`Send`,onChange:A}}]}),"props/required":()=>(0,k.jsx)(p,{label:`Output`,onRemove:A,rows:[z]}),"props/long":()=>(0,k.jsx)(p,{label:`Long`,onRemove:A,rows:[{key:`geo-location`,label:`{geo-location}`,code:!0,control:{type:`text`,value:`Hamilton, Waikato, New Zealand, Southern Hemisphere, Earth`,onChange:A}},{key:`averylongtokennamethatdoesnotbreak`,label:`{averylongtokennamethatdoesnotbreak}`,code:!0,control:{type:`text`,value:``,onChange:A}}]}),"props/cards":()=>(0,k.jsx)(He,{}),"props/widths":()=>(0,k.jsx)(p,{label:`Widths`,rows:[{key:`type`,label:`Type`,required:!0,control:{type:`select`,value:`ollama`,onChange:A,options:[{value:`ollama`,label:`Ollama (Local)`},{value:`anthropic`,label:`Anthropic`}]}},{key:`url`,label:`Address`,width:`full`,control:{type:`text`,value:``,placeholder:`http://127.0.0.1:11434`,onChange:A}},{key:`model`,label:`Model`,width:`full`,control:{type:`text`,value:`minicpm-v4.6:latest`,onChange:A}},{key:`seed`,label:`Seed`,control:{type:`text`,value:``,placeholder:`1234`,onChange:A}},{key:`nPredict`,label:`Reply tokens`,width:`quarter`,control:{type:`text`,value:``,onChange:A}},{key:`temperature`,label:`Temperature`,control:{type:`text`,value:`0.2`,onChange:A}},{key:`grammar`,label:`Grammar`,control:{type:`textarea`,value:`root ::= item (", " item)*
+item ::= [a-z ]+`,onChange:A}}]}),"props/affix":()=>(0,k.jsx)(p,{label:`Call`,rows:[{key:`name`,label:`Name`,control:{type:`text`,value:`open-meteo-forecast`,suffix:`.txt`,onChange:A}},{key:`taken`,label:`Name`,error:`open-meteo-forecast-storm.txt is taken.`,control:{type:`text`,value:`open-meteo-forecast-storm`,suffix:`.txt`,onChange:A}},{key:`status`,label:`Status`,width:`quarter`,control:{type:`text`,value:`200`,numeric:!0,onChange:A}},{key:`body`,label:`Body`,control:{type:`textarea`,mono:!0,value:`{
+  "category": "utilities"
+}`,onChange:A}}]}),"props/secret":()=>(0,k.jsx)(p,{label:`Keys`,rows:[{key:`held`,label:`Key`,control:{type:`secret`,value:`sk-not-a-real-key-0000`,onChange:A}},{key:`empty`,label:`Key`,control:{type:`secret`,value:``,onChange:A}},{key:`server`,label:`Key`,control:{type:`secret`,value:``,held:!0,onChange:A}}]}),"props/code":()=>(0,k.jsx)(p,{label:`Debug`,rows:[{key:`endpoint`,label:`Endpoint`,required:!0,control:{type:`code`,text:`https://api.anthropic.com/v1/messages`}},{key:`prompt`,label:`Prompt`,required:!0,control:{type:`static`,text:`(prompt)`}},{key:`model`,label:`model`,control:{type:`code`,text:`claude-sonnet-4-5-20250929`}},{key:`options`,label:`options`,control:{type:`code`,text:`{"seed":1234,"temperature":0}`}},{key:`stream`,label:`stream`,control:{type:`code`,text:`false`}}]}),"table/sortable":()=>(0,k.jsx)($,{count:4}),"table/paged":()=>(0,k.jsx)($,{count:35,pageSize:10}),"table/select":()=>(0,k.jsx)($,{count:5,picked:[]}),"table/select-on":()=>(0,k.jsx)($,{count:5,picked:[1,3]}),"table/filter":()=>(0,k.jsx)(Ve,{}),"table/filter-on":()=>(0,k.jsx)(Ve,{search:`IMG_40`,size:`large`}),"edit/form":()=>(0,k.jsx)(F,{outcome:`Connected`,menu:[{label:`Restore…`,icon:m.restore,onSelect:A},{label:`Apply preset: repeat 1.3`,icon:m.preset,onSelect:A}],sections:[{key:`identity`,rows:I()},{key:`settings`,title:`Settings`,rows:L()},{key:`debug`,title:`Debug`,rows:R(),folded:!0}]}),"edit/plain":()=>(0,k.jsx)(f,{title:`A mapping`,sections:[{key:`one`,rows:[z,{key:`with`,label:`Separator`,control:{type:`text`,value:`, `,onChange:A}}]}],onClose:A}),"edit/long":()=>(0,k.jsx)(F,{menu:[{label:`Restore…`,icon:m.restore,onSelect:A,disabled:!0}],sections:[{key:`identity`,rows:I()},{key:`settings`,title:`Settings`,rows:[...L(),{key:`grammar`,label:`Grammar`,control:{type:`textarea`,value:`root ::= item (", " item)*
+item ::= [a-z ]+`,onChange:A}},{key:`imagePosition`,label:`Image position`,control:{type:`select`,value:`after`,onChange:A,options:[{value:`before`,label:`before the prompt`},{value:`after`,label:`after the prompt`}]}},{key:`cachePrompt`,label:`Cache prompt`,control:{type:`select`,value:`off`,onChange:A,options:[{value:`on`,label:`on`},{value:`off`,label:`off`}]}}]},{key:`debug`,title:`Debug`,rows:R(),folded:!0}]}),"edit/status":()=>(0,k.jsx)(F,{status:`the relay could not reach http://127.0.0.1:11434`,statusBad:!0,sections:[{key:`identity`,rows:I()}]}),"view/tokens":()=>(0,k.jsx)(d,{id:`g`,title:`Token mappings`,level:3,children:(0,k.jsx)(Oe,{})}),"view/output":()=>(0,k.jsx)(N,{start:{kind:`list`,parse:`csv`,modifiers:[{type:`case`,mode:`LOWER`}]}}),"view/output-text":()=>(0,k.jsx)(N,{start:{kind:`text`,modifiers:[]}}),"view/output-two":()=>(0,k.jsx)(N,{start:{kind:`list`,parse:`csv`,modifiers:[{type:`case`,mode:`TITLE`},{type:`join`,with:` | `}]}}),"view/output-twice":()=>(0,k.jsx)(N,{start:{kind:`list`,parse:`csv`,modifiers:[{type:`case`,mode:`TITLE`},{type:`case`,mode:`LOWER`}]}}),"view/output-none":()=>(0,k.jsx)(N,{start:{kind:`text`,modifiers:[]},format:!1}),"view/output-flow":()=>(0,k.jsx)(N,{start:{kind:`text`,modifiers:[]},readAs:`@body('Ask')?['content']?[0]?['text']`}),"view/field-chip":()=>(0,k.jsx)(d,{id:`g`,title:`Edited`,level:3,children:(0,k.jsxs)(`div`,{className:`prompt`,children:[(0,k.jsx)(`span`,{className:`prop-name`,children:`Prompt`}),(0,k.jsx)(w,{words:`Summarise @{triggerBody()?['text']} in two short sentences for @{variables('Audience')}.`,was:`Summarise @{triggerBody()?['text']} in one sentence for @{variables('Audience')}.`})]})}),"view/field-edit":()=>(0,k.jsx)(ze,{}),"view/insert-field":()=>(0,k.jsx)(d,{id:`g`,title:`Edited`,level:3,children:(0,k.jsxs)(`div`,{className:`prompt`,children:[(0,k.jsxs)(`div`,{className:`pe-head`,children:[(0,k.jsx)(`span`,{className:`prop-name`,children:`Prompt`}),(0,k.jsx)(ne,{fields:J,onInsert:A,label:`Edited`,defaultOpen:!0})]}),(0,k.jsx)(`textarea`,{"aria-label":`Edited, prompt`,rows:4,defaultValue:q}),(0,k.jsx)(w,{className:`pe-words`,words:q,was:K,showExprs:!1})]})}),"multi/closed":()=>(0,k.jsx)(Y,{}),"multi/open":()=>(0,k.jsx)(Y,{open:!0}),"chiplist/words":()=>(0,k.jsx)(X,{start:[`clearly`,`obviously`,`as an AI`]}),"chiplist/empty":()=>(0,k.jsx)(X,{start:[]}),"wizard/pill":()=>(0,k.jsx)(Z,{}),"wizard/pill-open":()=>(0,k.jsx)(Z,{open:!0}),"switch/off":()=>(0,k.jsx)(G,{}),"switch/on":()=>(0,k.jsx)(G,{start:!0}),"switch/disabled":()=>(0,k.jsx)(s,{checked:!0,disabled:!0,label:`Snapshot now`,onChange:A})};function G({start:e}){let[t,n]=(0,E.useState)(!!e);return(0,k.jsx)(s,{checked:t,onChange:n,label:`Snapshot now`})}var K=`Tell me about this location in 1 sentence: @{variables('LocationVar')}`,q=`Tell me about this location in 2 sentences, for someone visiting for the first time: @{variables('LocationVar')}`,J=[{name:`LocationVar`,expr:`@{variables('LocationVar')}`},{name:`text`,expr:`@{triggerBody()?['text']}`}];function ze(){let[e,t]=(0,E.useState)(q);return(0,k.jsx)(d,{id:`g`,title:`Edited`,level:3,children:(0,k.jsx)(me,{label:`Edited`,asks:[{key:`connection`,label:`Connection`,control:{type:`static`,text:`Anthropic Direct`}}],prompt:e,onPrompt:t,rows:[{key:`model`,label:`Model`,control:{type:`static`,text:`claude-haiku-4-5`}}],request:[{key:`request`,label:`Request`,control:{type:`code`,text:`POST /v1/messages`}},{key:`body`,label:`Body`,control:{type:`code`,text:JSON.stringify({model:`claude-haiku-4-5`,max_tokens:1024,messages:[{role:`user`,content:e}]},null,2)}}],requestId:`story-edited-request`,words:{text:e,was:K},fields:J})})}function Y({open:e=!1}){let[t,n]=(0,E.useState)([`better`,`same`]),r=[{value:`better`,label:`Better`},{value:`same`,label:`Same`},{value:`worse`,label:`Worse`}];return e?(0,k.jsx)(`div`,{style:{minHeight:180},children:(0,k.jsx)(l,{value:t,options:r,onChange:n,defaultOpen:!0})}):(0,k.jsx)(p,{label:`Judged against recorded`,rows:[{key:`expected`,label:`Expected`,required:!0,control:{type:`multi`,value:t,options:r,onChange:n}}]})}function X({start:e}){let[t,n]=(0,E.useState)(e);return(0,k.jsx)(`div`,{style:{maxWidth:360},children:(0,k.jsx)(Se,{value:t,onChange:n})})}function Z({open:e}){return(0,k.jsx)(`div`,{style:{display:`flex`,justifyContent:`flex-end`,padding:`var(--s3)`},children:(0,k.jsx)(c,{name:`Test a workflow`,done:2,total:4,current:2,steps:[{title:`Add a workflow Source`,done:!0},{title:`Test an action`,done:!0},{title:`Run the test`,done:!1},{title:`Compare the replies`,done:!1}],onOpen:A,onCancel:A,...e?{defaultOpen:!0}:{}})})}var Q=De.map(e=>({...e,render:Re[e.path]})),Be=e=>Array.from({length:e},(e,t)=>({name:`IMG_${String(4e3+t*7919%97).padStart(4,`0`)}.jpg`,kb:t*37%900+40}));function Ve({search:e=``,size:t=`all`}){let[n,r]=(0,E.useState)(e),[i,o]=(0,E.useState)(t),[s,c]=(0,E.useState)(`shoot`),l=Be(12),u=l.filter(e=>e.name.toLowerCase().includes(n.trim().toLowerCase())&&(i===`all`||i===`large`==e.kb>=400));return(0,k.jsxs)(d,{id:`files`,title:`Files`,children:[(0,k.jsx)(a,{noun:`files`,search:{label:`Name`,placeholder:`Search names`,value:n,onChange:r},count:{shown:u.length,total:l.length},filters:[{key:`source`,label:`Source`,value:s,neutral:null,onChange:c,options:[{value:`shoot`,label:`Tuesday shoot`},{value:`samples`,label:`Sample library`}]},{key:`size`,label:`Size`,value:i,onChange:o,options:[{value:`all`,label:`All`},{value:`small`,label:`Small`},{value:`large`,label:`Large`}]}]}),u.length?(0,k.jsx)(b,{label:`Files`,rows:u,rowKey:e=>e.name,columns:[{key:`name`,header:`Name`,cell:e=>e.name},{key:`size`,header:`Size`,sort:e=>e.kb,cell:e=>`${e.kb} KB`}]}):(0,k.jsx)(`p`,{className:`note`,children:`No files match these filters.`})]})}function $({count:e,pageSize:t,picked:n}){let r=Be(e),[i,a]=(0,E.useState)(r),[o,s]=(0,E.useState)(()=>new Set((n??[]).map(e=>r[e].name)));return(0,k.jsx)(d,{id:`files`,title:`Files`,children:(0,k.jsx)(b,{label:`Files`,cards:!0,rows:i,rowKey:e=>e.name,pageSize:t??null,select:n&&{noun:`file`,selected:o,onChange:s,actions:e=>[{label:`Move…`,icon:m.rename,onSelect:A},{label:`Remove`,icon:m.remove,destructive:!0,onSelect:()=>{a(t=>t.filter(t=>!e.includes(t))),s(new Set)}},{label:`Export`,icon:C,onSelect:A}]},columns:[{key:`name`,header:`Name`,role:`title`,cell:e=>e.name},{key:`size`,header:`Size`,className:`num`,sort:e=>e.kb,cell:e=>`${e.kb} KB`},{key:`menu`,header:``,role:`menu`,sort:null,cell:e=>(0,k.jsx)(u,{label:`${e.name} actions`,actions:[{label:`Rename…`,icon:m.rename,onSelect:A},{label:`Remove`,icon:m.remove,destructive:!0,onSelect:A}]})}]})})}function He(){let[e,t]=(0,E.useState)(`files`);return(0,k.jsx)(p,{label:`New Source`,rows:[{key:`type`,label:`Type`,control:{type:`cards`,value:e,onChange:t,options:[{value:`files`,label:`File Library`,icon:(0,k.jsx)(r,{size:16,fill:`currentColor`,"aria-hidden":`true`})},{value:`flow`,label:`Power Automate flow`,icon:(0,k.jsx)(ye,{size:16,fill:`currentColor`,"aria-hidden":`true`})}]}},{key:`name`,label:`Name`,control:{type:`text`,value:``,onChange:A}}]})}var Ue=()=>decodeURIComponent(location.hash.replace(/^#\/?/,``));function We(){let[e,t]=(0,E.useState)(Ue);(0,E.useEffect)(()=>{let e=()=>t(Ue());return window.addEventListener(`hashchange`,e),()=>window.removeEventListener(`hashchange`,e)},[]);let n=Q.find(t=>t.path===e);return n?(0,k.jsxs)(`main`,{className:`gallery-one`,"data-route":n.path,children:[(0,k.jsxs)(`nav`,{className:`gallery-nav`,children:[(0,k.jsx)(`a`,{href:`#/`,children:`All`}),(0,k.jsx)(`span`,{children:n.title})]}),n.render()]}):(0,k.jsxs)(`main`,{className:`gallery-index`,"data-route":``,children:[(0,k.jsx)(`h1`,{children:`Components`}),(0,k.jsx)(`ul`,{children:Q.map(e=>(0,k.jsxs)(`li`,{children:[(0,k.jsx)(`a`,{href:`#/${e.path}`,children:e.title}),` `,(0,k.jsx)(`code`,{children:e.path})]},e.path))})]})}(0,Te.createRoot)(document.getElementById(`root`)).render((0,k.jsx)(E.StrictMode,{children:(0,k.jsx)(We,{})}));