@ancha-labs/cli 0.1.0 → 0.1.1

package/lib/assets/ui/app.js

@@ -159,7 +159,7 @@ function seedData(scene, selected) {
   return `<h2>${h(toolName(selected))}</h2>${body || empty("No data seeded")}${details("Full starting data", s)}${selected === "datadog" ? details("Generated telemetry definitions", { metrics: scene.world_rules?.metrics, logs: scene.world_rules?.logs, monitors: scene.world_rules?.monitors }) : ""}`;
 }
 var countLabel = (count, noun, plural = noun + "s") => `${count} ${count === 1 ? noun : plural}`;
-var toolNames = { slack: "Slack", datadog: "Datadog", pagerduty: "PagerDuty", github: "GitHub", "linkedin-ads": "LinkedIn Ads", salesforce: "Salesforce", gong: "Gong", "google-ads": "Google Ads", "chatgpt-ads": "ChatGPT Ads", hubspot: "HubSpot", ga4: "Google Analytics" };
+var toolNames = { slack: "Slack", "open-meteo": "Open-Meteo", datadog: "Datadog", pagerduty: "PagerDuty", github: "GitHub", "linkedin-ads": "LinkedIn Ads", salesforce: "Salesforce", gong: "Gong", "google-ads": "Google Ads", "chatgpt-ads": "ChatGPT Ads", hubspot: "HubSpot", ga4: "Google Analytics" };
 var toolName = (name) => toolNames[name] ?? name;
 var toolLinks = (names) => names.map((name) => `<a class="tool-chip" href="/tools/${encodeURIComponent(name)}" data-nav>${h(toolName(name))}</a>`).join("");
 var toolDescription = (name) => ({ slack: "Channels, messages and threads", datadog: "Monitors, metrics and logs", pagerduty: "Incidents, on-call and escalation", github: "Repositories, pull requests and deployments" })[name] ?? "API capabilities";
@@ -325,240 +325,6 @@ function toolLogo(name) {
   return svg ? `<img class="catalog-brand" src="data:image/svg+xml,${encodeURIComponent(svg)}" alt="" width="36" height="36">` : '<span class="catalog-brand" aria-hidden="true">\u25C7</span>';
 }
 
-// frontend/src/world-workspace.ts
-var enc = encodeURIComponent;
-var label = (task, index) => task.title && task.title !== String(task.prompt).split("\n")[0]?.slice(0, 120) ? task.title : "Task " + (index + 1);
-function worldSetups(tests, worldId, agentId) {
-  return tests.filter((test) => !test.archived && test.worldId === worldId && (!agentId || test.agentId === agentId));
-}
-function setupRuns(runs, test, exactRevision = true) {
-  return runs.filter((run) => !run.entries?.length && run.test?.id === test.id && (!exactRevision || run.test.revision === test.revision) && run.agentId === test.agentId && run.world?.id === test.worldId && (!exactRevision || run.world.revision === test.worldRevision)).sort((a, b) => String(b.createdAt).localeCompare(String(a.createdAt)));
-}
-function setupHistory(runs, test) {
-  const singles = new Set(setupRuns(runs, test, false).map((run) => run.id));
-  return runs.flatMap((run) => {
-    if (singles.has(run.id)) return [{ run, attemptCount: run.attempts?.length ?? 0, earlier: run.test.revision !== test.revision || run.world.revision !== test.worldRevision, grouped: false }];
-    const entries = (run.entries ?? []).filter((entry) => entry.test?.id === test.id && entry.agentId === test.agentId && entry.world?.id === test.worldId && entry.test.agentId === test.agentId && entry.test.worldId === test.worldId && entry.test.worldRevision === entry.world.revision);
-    if (!entries.length) return [];
-    const ids = new Set(entries.map((entry) => entry.id));
-    return [{
-      run,
-      attemptCount: (run.attempts ?? []).filter((attempt) => ids.has(attempt.entryId)).length,
-      earlier: entries.some((entry) => entry.test.revision !== test.revision || entry.world.revision !== test.worldRevision),
-      grouped: true
-    }];
-  }).sort((a, b) => String(b.run.createdAt).localeCompare(String(a.run.createdAt)));
-}
-function worldHistory(runs, worldId, agentId) {
-  return runs.flatMap((run) => {
-    if (!run.entries?.length) return run.world?.id === worldId && (!agentId || run.agentId === agentId) ? [{ run, attemptCount: run.attempts?.length ?? 0, grouped: false }] : [];
-    const entries = run.entries.filter((entry) => entry.world?.id === worldId && (!agentId || entry.agentId === agentId));
-    if (!entries.length) return [];
-    const ids = new Set(entries.map((entry) => entry.id));
-    return [{ run, attemptCount: (run.attempts ?? []).filter((attempt) => ids.has(attempt.entryId)).length, grouped: true }];
-  }).sort((a, b) => String(b.run.createdAt).localeCompare(String(a.run.createdAt)));
-}
-function workspaceDefinition(draft) {
-  return {
-    name: draft.name,
-    agentId: draft.agentId,
-    worldId: draft.worldId,
-    worldRevision: Number(draft.worldRevision),
-    tasks: draft.tasks.map((task) => ({ id: task.id, ...String(task.title ?? "").trim() ? { title: String(task.title).trim() } : {}, prompt: task.prompt, attempts: Number(task.attempts), timeoutMs: Number(task.timeoutMs) })),
-    concurrency: Number(draft.concurrency),
-    seed: draft.seed ?? "playground-1"
-  };
-}
-function mountWorldWorkspace(node, options) {
-  const { world, agents, api: api2, go: go2 } = options, base = "/worlds/" + enc(world.id), q = new URLSearchParams(location.search);
-  let tests = [], runs = [], errors = [], selected, agentId = q.get("agent") ?? "", configure = q.has("configure"), showHistory = false, saving = false, disposed = false, error = "", loaded = false;
-  const revisions = /* @__PURE__ */ new Map([[world.revision, world]]);
-  let draft, timer;
-  const launchPending = () => {
-    if (!selected) return false;
-    try {
-      return sessionStorage.getItem("ancha-saved-test-run:" + selected.id) !== null;
-    } catch {
-      return true;
-    }
-  };
-  const key = () => "ancha-world-setup:" + world.id + ":" + (selected?.id ?? "new:" + agentId);
-  const remember = () => {
-    if (draft) try {
-      sessionStorage.setItem(key(), JSON.stringify(draft));
-    } catch {
-    }
-  };
-  function getDraft() {
-    if (draft) return draft;
-    try {
-      const stored = JSON.parse(sessionStorage.getItem(key()) ?? "null");
-      if (stored && Array.isArray(stored.tasks)) draft = stored;
-    } catch {
-    }
-    draft ??= selected ? { ...selected, tasks: structuredClone(selected.tasks) } : { name: world.name, agentId, worldId: world.id, worldRevision: world.revision, concurrency: 1, seed: "playground-1", tasks: [{ id: "task_" + crypto.randomUUID(), title: "", prompt: "", attempts: 1, timeoutMs: 3e5 }] };
-    return draft;
-  }
-  function capture() {
-    const form = node.querySelector(".ww-form");
-    if (!form) return;
-    const d = getDraft(), values = new FormData(form);
-    d.name = String(values.get("name") ?? "");
-    d.agentId = String(values.get("agentId") ?? "");
-    d.concurrency = Number(values.get("concurrency"));
-    d.worldRevision = Number(values.get("worldRevision"));
-    d.tasks = [...form.querySelectorAll(".ww-task-editor")].map((row) => ({ id: row.dataset.taskId, title: row.querySelector("[name=title]").value, prompt: row.querySelector("[name=prompt]").value, attempts: Number(row.querySelector("[name=attempts]").value), timeoutMs: Math.round(Number(row.querySelector("[name=minutes]").value) * 6e4) }));
-    remember();
-  }
-  function href2(test, agent = agentId, extra = "") {
-    return base + "?" + new URLSearchParams({ ...agent ? { agent } : {}, ...test ? { test: test.id } : {}, ...extra ? { [extra]: "1" } : {} });
-  }
-  function configuration2() {
-    const d = getDraft(), pinned = Number(d.worldRevision) !== world.revision, startingWorld = revisions.get(Number(d.worldRevision));
-    return `<section class="ww-configure"><div class="ww-section-heading"><h2>Configure</h2><button type="button" data-ww="close">Close</button></div><form class="ww-form"><div class="form-grid"><label>Agent<select name="agentId" required><option value="">Choose agent</option>${agents.map((a) => `<option value="${h(a.id)}" ${d.agentId === a.id ? "selected" : ""}>${h(a.name)}</option>`).join("")}</select></label><label>Configuration name<input name="name" required maxlength="120" value="${h(d.name)}"></label></div><section class="ww-tools"><div class="ww-section-heading"><h3>Tools</h3><a data-nav href="${base}?edit=1">Edit world</a></div><div class="ww-tool-list">${(startingWorld?.configuration.preset.datasets ?? []).map((item) => `<a data-nav href="/tools/${enc(item.tool)}">${toolLogo(item.tool)}${h(toolName(item.tool))}</a>`).join("")}</div>${pinned ? `<label>Starting data<select name="worldRevision"><option value="${h(d.worldRevision)}">Saved revision ${h(d.worldRevision)}</option><option value="${world.revision}">Current revision ${world.revision}</option></select></label>` : `<input type="hidden" name="worldRevision" value="${h(d.worldRevision)}">`}</section><div class="ww-section-heading"><h3>Tasks</h3><button type="button" data-ww="add-task" ${d.tasks.length >= 20 ? "disabled" : ""}>Add task</button></div><div class="ww-task-fields">${d.tasks.map((task, index) => `<article class="ww-task-editor" data-task-id="${h(task.id)}"><div class="ww-section-heading"><h3>Task ${index + 1}</h3>${d.tasks.length > 1 ? `<button type="button" data-ww="remove-task" data-value="${h(task.id)}">Remove</button>` : ""}</div><label>Task name <span class="muted">optional</span><input name="title" maxlength="120" value="${h(task.title ?? "")}" placeholder="Task ${index + 1}"></label><label>Instructions<textarea name="prompt" required maxlength="20000" rows="4">${h(task.prompt)}</textarea></label><div class="ww-task-options"><label>Repeat<input type="number" name="attempts" min="1" max="100" value="${h(task.attempts)}" required><span>times</span></label><label>Time limit <span class="muted">minutes</span><input type="number" name="minutes" min="0.016666666666666666" max="60" step="any" value="${h(task.timeoutMs / 6e4)}" required></label></div></article>`).join("")}</div><div class="ww-footer"><label>Run at once<input name="concurrency" type="number" min="1" max="10" value="${h(d.concurrency)}" required></label><button type="submit" class="primary" ${saving ? "disabled" : ""}>${saving ? "Saving\u2026" : "Save configuration"}</button></div></form></section>`;
-  }
-  function render() {
-    if (disposed) return;
-    const currentRuns = selected ? setupRuns(runs, selected) : [], run = currentRuns[0], attempts = run?.attempts ?? [], tasks = selected?.tasks ?? [], choices = worldSetups(tests, world.id, agentId);
-    node.innerHTML = `<div class="ww-workspace"><header class="ww-header"><h1>${h(world.name)}</h1><div class="ww-world-actions"><label class="ww-agent-select">Agent<select data-ww-select="agent"><option value="" disabled>Choose agent</option>${agents.map((a) => `<option value="${h(a.id)}" ${a.id === agentId ? "selected" : ""}>${h(a.name)}</option>`).join("")}</select></label><button type="button" data-ww="configure">Configure</button>${selected ? options.runActions?.(selected.id) ?? button("Run tasks", "run-saved-test", selected.id, "primary") : ""}</div></header><p class="ww-error" role="alert" ${error ? "" : "hidden"}>${h(error)}</p>${errors.map((e) => `<p class="ww-error" role="alert">${h(e.message)}</p>`).join("")}${!loaded ? '<p role="status">Loading tasks\u2026</p>' : ""}${choices.length > 1 ? `<label class="ww-setup-select">Configuration<select data-ww-select="test">${!selected ? '<option value="" selected disabled>New configuration</option>' : ""}${choices.map((t) => `<option value="${h(t.id)}" ${t.id === selected?.id ? "selected" : ""}>${h(t.name)}</option>`).join("")}</select></label>` : ""}${loaded && configure ? configuration2() : loaded ? `<section class="ww-summary-widgets"><article><span>Success rate</span><strong class="muted">\u2014</strong></article><article><span>Attempts</span><strong>${attempts.length}</strong></article><article><span>Ungraded</span><strong>${attempts.length}</strong></article></section><div class="ww-section-heading"><h2>Tasks${tasks.length ? ` <span class="ww-task-count">${tasks.length}</span>` : ""}</h2><div class="actions">${worldHistory(runs, world.id, agentId || void 0).length || selected ? `<button type="button" data-ww="history" aria-expanded="${showHistory}">History</button>` : ""}<button type="button" data-ww="configure">${tasks.length ? "Edit tasks" : "Add tasks"}</button></div></div>${tasks.length ? `<section class="ww-task-table"><div class="ww-table-head"><span>Task</span><span>Success rate</span><span>Attempts</span><span></span></div>${tasks.map((task, index) => {
-      const list2 = attempts.filter((a) => a.taskId === task.id), taskURL = run ? `/playgrounds/${enc(run.id)}?task=${enc(task.id)}${list2[0] ? "&attempt=" + enc(list2[0].id) : ""}` : "";
-      return `<article class="ww-task-row"><div class="ww-task-identity">${taskURL ? `<a data-nav class="ww-task-name" href="${taskURL}"><span class="ww-ordinal">${index + 1}</span><strong>${h(label(task, index))}</strong></a>` : `<strong class="ww-task-name"><span class="ww-ordinal">${index + 1}</span>${h(label(task, index))}</strong>`}<p class="ww-task-prompt" title="${h(task.prompt)}">${h(task.prompt)}</p></div><div class="ww-rate">${list2.length ? "Ungraded" : "\u2014"}</div><div class="ww-attempt-marks" role="group" aria-label="Attempts for ${h(label(task, index))}">${list2.map((a) => `<a data-nav class="ww-attempt-mark ${h(a.status)}" href="/playgrounds/${enc(run.id)}?task=${enc(task.id)}&attempt=${enc(a.id)}" title="Attempt ${a.number}: ${h(attemptLabel(a.status))} \xB7 Ungraded" aria-label="Attempt ${a.number}: ${h(attemptLabel(a.status))} \xB7 Ungraded">${a.status === "error" ? "!" : ["running", "starting"].includes(a.status) ? "\u25CC" : "\u2212"}</a>`).join("") || `<span class="muted">Repeat ${h(task.attempts)} times</span>`}</div>${taskURL ? `<a data-nav class="ww-row-inspect" href="${taskURL}" aria-label="Inspect ${h(label(task, index))}">\u2192</a>` : "<span></span>"}</article>`;
-    }).join("")}</section>` : `<div class="quiet-empty">${agents.length ? "Add tasks to this world." : '<a data-nav href="/agents">Connect an agent</a>'}</div>`}${selected && !currentRuns.length && setupHistory(runs, selected).some((item) => item.earlier) ? '<p class="ww-snapshot-notice">This configuration has changed. Earlier attempts are in History.</p>' : ""}${showHistory ? history2() : ""}` : ""}</div>`;
-  }
-  function history2() {
-    const all = worldHistory(runs, world.id, agentId || void 0);
-    return `<section class="ww-history"><div class="ww-section-heading"><h2>History</h2><button type="button" data-ww="new">New configuration</button></div>${all.length ? all.map(({ run, attemptCount, grouped }) => `<a data-nav class="ww-history-row" href="/playgrounds/${enc(run.id)}"><time>${h(new Date(run.createdAt).toLocaleString())}</time><span>${h(attemptLabel(run.status))}</span><span>${attemptCount} attempts</span>${grouped ? "<span>Group</span>" : ""}${run.test?.name || run.groupName ? `<span>${h(run.groupName ?? run.test.name)}</span>` : ""}<span>\u2192</span></a>`).join("") : '<p class="quiet-empty">No attempts yet.</p>'}</section>`;
-  }
-  async function refresh(initial = false) {
-    try {
-      const [definitions, batches] = await Promise.all([api2("/playground-tests"), api2("/playgrounds")]);
-      if (disposed) return;
-      tests = definitions.tests ?? [];
-      runs = batches.runs ?? [];
-      errors = [...definitions.errors ?? [], ...batches.errors ?? []];
-      if (initial) {
-        const requested = q.get("test");
-        if (requested) {
-          selected = worldSetups(tests, world.id).find((test) => test.id === requested);
-          if (!selected || agentId && selected.agentId !== agentId) {
-            selected = void 0;
-            error = "This configuration is unavailable for this world and agent.";
-          } else agentId = selected.agentId;
-        } else if (q.has("new")) {
-          selected = void 0;
-          agentId ||= agents[0]?.id ?? "";
-        } else {
-          selected = worldSetups(tests, world.id, agentId || void 0)[0];
-          const historical = worldHistory(runs, world.id)[0]?.run;
-          agentId ||= selected?.agentId ?? historical?.entries?.find((entry) => entry.world?.id === world.id)?.agentId ?? historical?.agentId ?? agents[0]?.id ?? "";
-        }
-      } else if (selected) selected = tests.find((test) => test.id === selected.id);
-      if (initial && !selected && worldHistory(runs, world.id, agentId || void 0).length) showHistory = true;
-      if (configure) getDraft();
-      options.onContext?.(selected);
-      const pinnedRevision = Number(draft?.worldRevision ?? selected?.worldRevision ?? world.revision);
-      if (!revisions.has(pinnedRevision)) {
-        const savedWorld = await api2(`/worlds/${enc(world.id)}?revision=${pinnedRevision}`);
-        if (disposed) return;
-        revisions.set(pinnedRevision, savedWorld);
-      }
-      loaded = true;
-      if (!configure) render();
-      else if (initial) render();
-      if (runs.some((run) => !terminal(run.status))) timer = setTimeout(() => {
-        void refresh();
-      }, 2e3);
-    } catch (e) {
-      if (!disposed) {
-        error = e instanceof Error ? e.message : String(e);
-        loaded = true;
-        render();
-      }
-    }
-  }
-  node.addEventListener("input", capture);
-  node.addEventListener("change", (event) => {
-    const control = event.target;
-    if (control.dataset.wwSelect === "agent") {
-      capture();
-      const match = worldSetups(tests, world.id, control.value)[0];
-      go2(href2(match, control.value));
-    } else if (control.dataset.wwSelect === "test") go2(href2(tests.find((t) => t.id === control.value)));
-    else {
-      capture();
-      if (control.name === "worldRevision") render();
-    }
-  });
-  node.addEventListener("click", (event) => {
-    const control = event.target.closest("[data-ww]");
-    if (!control) return;
-    event.preventDefault();
-    event.stopPropagation();
-    if (saving) return;
-    capture();
-    error = "";
-    switch (control.dataset.ww) {
-      case "configure":
-        if (launchPending()) {
-          error = "Resolve the pending run before editing this configuration.";
-          break;
-        }
-        configure = true;
-        break;
-      case "close":
-        configure = false;
-        break;
-      case "history":
-        showHistory = !showHistory;
-        break;
-      case "new":
-        go2(href2(void 0, agentId, "new") + "&configure=1");
-        return;
-      case "add-task":
-        if (getDraft().tasks.length < 20) getDraft().tasks.push({ id: "task_" + crypto.randomUUID(), title: "", prompt: "", attempts: 1, timeoutMs: 3e5 });
-        break;
-      case "remove-task":
-        if (getDraft().tasks.length > 1) getDraft().tasks = getDraft().tasks.filter((t) => t.id !== control.dataset.value);
-        break;
-    }
-    remember();
-    render();
-  });
-  node.addEventListener("submit", (event) => {
-    const form = event.target;
-    if (!form.matches(".ww-form")) return;
-    event.preventDefault();
-    event.stopPropagation();
-    if (saving) return;
-    if (launchPending()) {
-      error = "Resolve the pending run before editing this configuration.";
-      render();
-      return;
-    }
-    capture();
-    saving = true;
-    error = "";
-    render();
-    const d = getDraft(), body = workspaceDefinition(d), oldKey = key();
-    void api2(selected ? `/playground-tests/${enc(selected.id)}/update` : "/playground-tests", { ...body, ...selected ? { revision: d.revision } : {} }).then((saved) => {
-      sessionStorage.removeItem(oldKey);
-      if (!disposed) go2(href2(saved, saved.agentId));
-    }).catch((e) => {
-      if (!disposed) {
-        error = e instanceof Error ? e.message : String(e);
-        saving = false;
-        render();
-      }
-    });
-  });
-  render();
-  void refresh(true);
-  return () => {
-    disposed = true;
-    clearTimeout(timer);
-  };
-}
-
 // frontend/src/experience.ts
 var list = (value) => Array.isArray(value) ? value.filter((v) => v && typeof v === "object") : [];
 var text = (value) => typeof value === "string" ? value : typeof value === "number" || typeof value === "boolean" ? String(value) : "";
@@ -915,7 +681,15 @@ function taskDisplayName(task, index) {
   const title2 = typeof task.title === "string" ? task.title.trim().replace(/\s+/g, " ") : "";
   return title2 && title2.length <= 80 ? title2 : `Task ${index + 1}`;
 }
+function attemptOutcome(attempt) {
+  if (attempt.outcome === "passed") return { key: "passed", label: "Passed", tone: "pass", mark: "\u2713" };
+  if (attempt.outcome === "failed") return { key: "failed", label: "Failed", tone: "fail", mark: "\u2717" };
+  return { key: "ungraded", label: "Ungraded", tone: "neutral", mark: "\u2013" };
+}
 function executionSummary(attempts) {
+  const finished = attempts.filter((a) => ["completed", "error", "stopped"].includes(a.status));
+  const durations = finished.map((a) => Date.parse(a.finishedAt) - Date.parse(a.startedAt)).filter((ms) => Number.isFinite(ms) && ms >= 0).sort((x, y) => x - y);
+  const passed = attempts.filter((a) => a.outcome === "passed").length, failed = attempts.filter((a) => a.outcome === "failed").length;
   return {
     total: attempts.length,
     completed: attempts.filter((a) => a.status === "completed").length,
@@ -923,17 +697,61 @@ function executionSummary(attempts) {
     queued: attempts.filter((a) => a.status === "queued").length,
     error: attempts.filter((a) => a.status === "error").length,
     stopped: attempts.filter((a) => a.status === "stopped").length,
-    ungraded: attempts.length
+    finished: finished.length,
+    passed,
+    failed,
+    graded: passed + failed,
+    ungraded: attempts.length - passed - failed,
+    toolCalls: attempts.reduce((sum, a) => sum + (Number(a.toolCalls) || 0), 0),
+    medianMs: durations.length ? durations[Math.floor((durations.length - 1) / 2)] : void 0
   };
 }
+var percent = (part, whole) => whole ? `${Math.round(part / whole * 100)}%` : "\u2014";
+var duration = (ms) => ms === void 0 ? "\u2014" : ms < 1e4 ? `${(ms / 1e3).toFixed(1)}s` : ms < 12e4 ? `${Math.round(ms / 1e3)}s` : `${Math.round(ms / 6e4)}m`;
+function summaryWidgets(attempts, wrapper = "aw-widgets", item = 'section class="aw-widget"') {
+  const s = executionSummary(attempts), tag = item.split(" ")[0];
+  const cell = (label3, value, note) => `<${item}><span>${h(label3)}</span><strong>${h(value)}</strong><small>${h(note)}</small></${tag}>`;
+  const progress = s.active || s.queued ? [s.active && `${s.active} running`, s.queued && `${s.queued} queued`].filter(Boolean).join(" \xB7 ") : s.error ? `${s.error} ended with an error` : s.total ? "All finished" : "No attempts yet";
+  const rate = s.graded ? `${s.passed} of ${s.graded} passed` : s.total ? "Add checks to a task to grade it" : "No attempts yet";
+  return `<div class="${wrapper} four">${cell("Success rate", percent(s.passed, s.graded), rate)}${cell("Finished", `${s.finished}/${s.total}`, progress)}${cell("Tool calls", String(s.toolCalls), s.finished ? `${(s.toolCalls / s.finished).toFixed(1)} per attempt` : "\u2014")}${cell("Median time", duration(s.medianMs), "per attempt")}</div>${progressBar(attempts)}`;
+}
+function attemptState(attempt) {
+  const outcome = attemptOutcome(attempt);
+  if (outcome.key !== "ungraded") return outcome.key;
+  return ["queued", "starting", "running", "completed", "stopped", "error"].includes(attempt.status) ? attempt.status === "starting" ? "running" : attempt.status : "queued";
+}
+function progressBar(attempts) {
+  if (!attempts.length) return "";
+  const s = executionSummary(attempts), label3 = `${s.finished} of ${s.total} attempts finished${s.graded ? `, ${s.passed} passed, ${s.failed} failed` : ""}`;
+  return `<div class="run-progress" role="img" aria-label="${h(label3)}" title="${h(label3)}">${attempts.map((a) => `<span class="seg ${attemptState(a)}"></span>`).join("")}</div>`;
+}
+function taskOutcome(attempts) {
+  const s = executionSummary(attempts);
+  if (s.graded) return { value: percent(s.passed, s.graded), note: `${s.passed} of ${s.graded} passed`, tone: s.passed === s.graded ? "pass" : s.passed ? "mixed" : "fail" };
+  if (!s.total) return { value: "\u2014", note: "No attempts", tone: "neutral" };
+  return { value: `${s.finished}/${s.total}`, note: s.active || s.queued ? "running" : "finished \xB7 no checks", tone: "neutral" };
+}
 function executionBadge(attempt) {
   const status = String(attempt.status ?? "unavailable");
   const tone = status === "error" ? "error" : ["running", "starting"].includes(status) ? "active" : "neutral";
   return `<span class="aw-execution ${tone}">${h(attemptLabel(status))}</span>`;
 }
+function outcomeBadge(attempt) {
+  const outcome = attemptOutcome(attempt);
+  return `<span class="aw-outcome ${outcome.tone}">${h(outcome.label)}</span>`;
+}
+function railMark(attempt) {
+  const outcome = attemptOutcome(attempt);
+  return `<span class="aw-rail-mark ${outcome.tone}" title="${h(outcome.label)}" aria-label="${h(outcome.label)}">${outcome.mark}</span>`;
+}
+function checksView(attempt) {
+  const checks2 = records(attempt.checks);
+  if (!checks2.length) return "";
+  const passed = checks2.filter((c) => c.passed).length;
+  return `<section class="aw-checks"><h3>Checks <span>${passed} of ${checks2.length} passed</span></h3><ul>${checks2.map((c) => `<li class="${c.passed ? "pass" : "fail"}"><span class="aw-check-mark" aria-hidden="true">${c.passed ? "\u2713" : "\u2717"}</span><div><strong>${h(c.id)}</strong><p>${h(c.expected)}</p><p class="muted">${h(c.error ? `Error: ${c.error}` : c.actual)}</p></div></li>`).join("")}</ul></section>`;
+}
 function widgets(attempts) {
-  const summary = executionSummary(attempts);
-  return `<div class="aw-widgets"><section class="aw-widget"><span>Success rate</span><strong>\u2014</strong></section><section class="aw-widget"><span>Attempts</span><strong>${summary.total}</strong></section><section class="aw-widget"><span>Ungraded</span><strong>${summary.ungraded}</strong></section></div>`;
+  return summaryWidgets(attempts);
 }
 function errorMessage(value) {
   if (!value) return "";
@@ -947,8 +765,9 @@ function entrySections(batch) {
 function markers(batch, task, attempts) {
   return `<div class="aw-attempt-strip" aria-label="${h(task.title || "Task")} attempts">${attempts.map((attempt) => {
     const state = ["queued", "starting", "running", "completed", "stopped", "error"].includes(attempt.status) ? attempt.status : "unavailable";
-    const label3 = `Attempt ${attempt.number} \xB7 ${attemptLabel(state)} \xB7 Ungraded`;
-    return `<a data-nav class="aw-attempt-marker execution-${h(state)}" href="${h(taskHref(batch, task, { attempt: attempt.id }))}" title="${h(label3)}" aria-label="${h(label3)}"><span>${h(attempt.number)}</span></a>`;
+    const outcome = attemptOutcome(attempt);
+    const label3 = `Attempt ${attempt.number} \xB7 ${attemptLabel(state)} \xB7 ${outcome.label}`;
+    return `<a data-nav class="aw-attempt-marker execution-${h(state)} outcome-${outcome.key}" href="${h(taskHref(batch, task, { attempt: attempt.id }))}" title="${h(label3)}" aria-label="${h(label3)}"><span>${h(attempt.number)}</span></a>`;
   }).join("")}</div>`;
 }
 function comparisonForm(batch, task, attempts, selected) {
@@ -964,8 +783,8 @@ function renderRunOverview(batch, actionsHtml = "", options = {}) {
     const shown = selected ? tasks.filter((t) => t.id === selected.id) : tasks;
     if (!shown.length) return "";
     return `<section class="aw-task-section">${groups.length > 1 ? `<header class="aw-entry-header"><h2>${h(entry?.agentName ?? entry?.name ?? "Agent")}</h2><span>${h(entry?.world?.name ?? entry?.test?.world?.name ?? "")}</span></header>` : ""}<div class="aw-section-heading"><h2>${selected ? "Attempts" : "Tasks"}</h2>${!selected ? `<span>${tasks.length}</span>` : ""}</div>${selected ? `<section class="aw-task-description"><p>${h(selected.prompt)}</p></section>` : ""}<div class="aw-task-table">${shown.map((task) => {
-      const taskAttempts = attempts.filter((a) => a.taskId === task.id), index = tasks.indexOf(task), first = taskAttempts[0];
-      return `<article class="aw-task-row"><div class="aw-task-name">${navLink(taskDisplayName(task, index), taskHref(batch, task, first ? { attempt: first.id } : {}))}${!selected ? `<p class="aw-prompt-excerpt">${h(task.prompt)}</p>` : ""}</div><div class="aw-task-outcome"><strong>\u2014</strong><span>Ungraded</span></div>${markers(batch, task, taskAttempts)}${taskAttempts.length ? "" : '<p class="quiet-empty">No attempts recorded.</p>'}${comparisonForm(batch, task, taskAttempts)}</article>`;
+      const taskAttempts = attempts.filter((a) => a.taskId === task.id), index = tasks.indexOf(task), first = taskAttempts[0], outcome = taskOutcome(taskAttempts);
+      return `<article class="aw-task-row"><div class="aw-task-name">${navLink(taskDisplayName(task, index), taskHref(batch, task, first ? { attempt: first.id } : {}))}${!selected ? `<p class="aw-prompt-excerpt">${h(task.prompt)}</p>` : ""}</div><div class="aw-task-outcome ${outcome.tone}"><strong>${h(outcome.value)}</strong><span>${h(outcome.note)}</span></div>${markers(batch, task, taskAttempts)}${taskAttempts.length ? "" : '<p class="quiet-empty">No attempts recorded.</p>'}${comparisonForm(batch, task, taskAttempts)}</article>`;
     }).join("")}</div></section>`;
   }).join("") || '<p class="quiet-empty">No tasks recorded.</p>'}</div>`;
 }
@@ -1082,7 +901,7 @@ function renderAttemptWorkspace(batch, attempt, evidence, options = {}) {
   const query = options.query ?? {};
   const toolsNav = `<nav class="tool-nav" aria-label="Attempt tools">${tools.map((t) => navLink(toolName(t), taskHref(batch, task, { attempt: attempt.id, view: "tools", tool: t }), t === tool ? "active" : "")).join("")}</nav>`;
   const content = view === "tools" ? toolsNav + (tool ? capturedToolView(observation2, tool, { ...query, baseHref: attemptBase }) : '<p class="quiet-empty">No captured tool data.</p>') : managedResponse(fullEvidence) + activityView(observation2, { baseHref: attemptBase });
-  return `<div class="aw-task-detail">${navLink("\u2190 Tasks", basePath(batch), "back")}<header class="aw-header"><h1>${h(taskDisplayName(task, Math.max(index, 0)))}</h1><div class="aw-actions">${options.actionsHtml ?? ""}</div></header><section class="aw-task-description"><p>${h(task.prompt)}</p></section>${errorMessage(batch.error)}<div class="aw-workspace"><aside class="aw-attempt-rail"><h2>Attempts</h2><nav class="aw-attempt-list" aria-label="Task attempts">${attempts.map((a) => `<a data-nav class="aw-attempt-item ${a.id === attempt.id ? "selected" : ""}" ${a.id === attempt.id ? 'aria-current="page"' : ""} href="${h(taskHref(batch, task, { attempt: a.id, view }))}"><span class="aw-attempt-number">Attempt ${h(a.number)}</span>${executionBadge(a)}</a>`).join("")}</nav></aside><section class="aw-evidence"><header class="aw-evidence-header"><div><h2>Attempt ${h(attempt.number)}</h2>${executionBadge(attempt)}<span class="aw-outcome">Ungraded</span></div>${comparisonForm(batch, task, attempts, attempt)}</header><nav class="aw-tabs" aria-label="Attempt evidence">${navLink("Activity", attemptBase, view === "activity" ? "active" : "").replace("<a ", `<a ${view === "activity" ? 'aria-current="page" ' : ""}`)}${navLink("Tools", taskHref(batch, task, { attempt: attempt.id, view: "tools" }), view === "tools" ? "active" : "").replace("<a ", `<a ${view === "tools" ? 'aria-current="page" ' : ""}`)}</nav><div id="attempt-evidence" class="aw-evidence-body">${errorMessage(fullEvidence.error)}${content}</div></section></div></div>`;
+  return `<div class="aw-task-detail">${navLink("\u2190 Tasks", basePath(batch), "back")}<header class="aw-header"><h1>${h(taskDisplayName(task, Math.max(index, 0)))}</h1><div class="aw-actions">${options.actionsHtml ?? ""}</div></header><section class="aw-task-description"><p>${h(task.prompt)}</p></section>${errorMessage(batch.error)}<div class="aw-workspace"><aside class="aw-attempt-rail"><h2>Attempts</h2><nav class="aw-attempt-list" aria-label="Task attempts">${attempts.map((a) => `<a data-nav class="aw-attempt-item ${a.id === attempt.id ? "selected" : ""}" ${a.id === attempt.id ? 'aria-current="page"' : ""} href="${h(taskHref(batch, task, { attempt: a.id, view }))}"><span class="aw-attempt-number">Attempt ${h(a.number)}</span>${executionBadge(a)}${a.outcome ? railMark(a) : ""}</a>`).join("")}</nav></aside><section class="aw-evidence"><header class="aw-evidence-header"><div><h2>Attempt ${h(attempt.number)}</h2>${executionBadge(attempt)}${outcomeBadge(attempt)}</div>${comparisonForm(batch, task, attempts, attempt)}</header>${checksView(attempt)}<nav class="aw-tabs" aria-label="Attempt evidence">${navLink("Activity", attemptBase, view === "activity" ? "active" : "").replace("<a ", `<a ${view === "activity" ? 'aria-current="page" ' : ""}`)}${navLink("Tools", taskHref(batch, task, { attempt: attempt.id, view: "tools" }), view === "tools" ? "active" : "").replace("<a ", `<a ${view === "tools" ? 'aria-current="page" ' : ""}`)}</nav><div id="attempt-evidence" class="aw-evidence-body">${errorMessage(fullEvidence.error)}${content}</div></section></div></div>`;
 }
 function renderManagedComparison(data) {
   const sides = records(data.attempts ?? [data.left, data.right]);
@@ -1091,7 +910,279 @@ function renderManagedComparison(data) {
   const summary = data.stateComparison;
   const stateSummary = !summary?.available ? "Final tool data is unavailable for one or both attempts." : !records(summary.differences).length ? "Same final tool data" : `Different final data: ${[...new Set(records(summary.differences).map((d) => String(d.path).split("/")[1]).filter(Boolean))].map(toolName).join(" \xB7 ")}${summary.truncated ? " \xB7 More differences in saved evidence" : ""}`;
   const cells = (render, extra = "") => sides.map((side) => `<section class="aw-compare-cell ${extra}">${render(side)}</section>`).join("");
-  return `<div class="aw-managed-comparison"><p class="aw-comparison-summary">${h(stateSummary)}</p><div class="aw-comparison">${cells((side) => `<h2>Attempt ${h(side.number)}</h2>${executionBadge(side)}<span class="aw-outcome">Ungraded</span>${navLink("Open attempt", url("/playgrounds/" + encodeURIComponent(String(data.runId)), { task: data.taskId ?? side.taskId, attempt: side.id }))}`, "aw-compare-head")}${cells((side) => errorMessage(side.error) + managedResponse(side), "aw-compare-response")}${cells((side) => activityView({ ...side, trace: side.toolActivity ?? side.trace ?? [], state: side.state ?? {} }, { baseHref: url("/playgrounds/" + encodeURIComponent(String(data.runId)), { task: data.taskId ?? side.taskId, attempt: side.id }) }), "aw-compare-activity")}</div></div>`;
+  return `<div class="aw-managed-comparison"><p class="aw-comparison-summary">${h(stateSummary)}</p><div class="aw-comparison">${cells((side) => `<h2>Attempt ${h(side.number)}</h2>${executionBadge(side)}${outcomeBadge(side)}${navLink("Open attempt", url("/playgrounds/" + encodeURIComponent(String(data.runId)), { task: data.taskId ?? side.taskId, attempt: side.id }))}`, "aw-compare-head")}${cells((side) => errorMessage(side.error) + managedResponse(side), "aw-compare-response")}${cells((side) => activityView({ ...side, trace: side.toolActivity ?? side.trace ?? [], state: side.state ?? {} }, { baseHref: url("/playgrounds/" + encodeURIComponent(String(data.runId)), { task: data.taskId ?? side.taskId, attempt: side.id }) }), "aw-compare-activity")}</div></div>`;
+}
+
+// frontend/src/world-workspace.ts
+var enc = encodeURIComponent;
+var label = (task, index) => task.title && task.title !== String(task.prompt).split("\n")[0]?.slice(0, 120) ? task.title : "Task " + (index + 1);
+function worldSetups(tests, worldId, agentId) {
+  return tests.filter((test) => !test.archived && test.worldId === worldId && (!agentId || test.agentId === agentId));
+}
+function setupRuns(runs, test, exactRevision = true) {
+  return runs.filter((run) => !run.entries?.length && run.test?.id === test.id && (!exactRevision || run.test.revision === test.revision) && run.agentId === test.agentId && run.world?.id === test.worldId && (!exactRevision || run.world.revision === test.worldRevision)).sort((a, b) => String(b.createdAt).localeCompare(String(a.createdAt)));
+}
+function setupHistory(runs, test) {
+  const singles = new Set(setupRuns(runs, test, false).map((run) => run.id));
+  return runs.flatMap((run) => {
+    if (singles.has(run.id)) return [{ run, attemptCount: run.attempts?.length ?? 0, earlier: run.test.revision !== test.revision || run.world.revision !== test.worldRevision, grouped: false }];
+    const entries = (run.entries ?? []).filter((entry) => entry.test?.id === test.id && entry.agentId === test.agentId && entry.world?.id === test.worldId && entry.test.agentId === test.agentId && entry.test.worldId === test.worldId && entry.test.worldRevision === entry.world.revision);
+    if (!entries.length) return [];
+    const ids = new Set(entries.map((entry) => entry.id));
+    return [{
+      run,
+      attemptCount: (run.attempts ?? []).filter((attempt) => ids.has(attempt.entryId)).length,
+      earlier: entries.some((entry) => entry.test.revision !== test.revision || entry.world.revision !== test.worldRevision),
+      grouped: true
+    }];
+  }).sort((a, b) => String(b.run.createdAt).localeCompare(String(a.run.createdAt)));
+}
+function worldHistory(runs, worldId, agentId) {
+  return runs.flatMap((run) => {
+    if (!run.entries?.length) return run.world?.id === worldId && (!agentId || run.agentId === agentId) ? [{ run, attemptCount: run.attempts?.length ?? 0, grouped: false }] : [];
+    const entries = run.entries.filter((entry) => entry.world?.id === worldId && (!agentId || entry.agentId === agentId));
+    if (!entries.length) return [];
+    const ids = new Set(entries.map((entry) => entry.id));
+    return [{ run, attemptCount: (run.attempts ?? []).filter((attempt) => ids.has(attempt.entryId)).length, grouped: true }];
+  }).sort((a, b) => String(b.run.createdAt).localeCompare(String(a.run.createdAt)));
+}
+function checksText(task) {
+  if (typeof task.checksText === "string") return task.checksText;
+  return Array.isArray(task.checks) && task.checks.length ? `[
+${task.checks.map((check) => "  " + JSON.stringify(check)).join(",\n")}
+]` : "";
+}
+function taskChecks(task, index) {
+  if (typeof task.checksText !== "string") return Array.isArray(task.checks) && task.checks.length ? task.checks : void 0;
+  const text2 = task.checksText.trim();
+  if (!text2) return void 0;
+  let parsed;
+  try {
+    parsed = JSON.parse(text2);
+  } catch {
+    throw new Error(`Task ${index + 1}: checks must be valid JSON, a list like [{"id": "posted", "trace": {"called": "POST /api/chat.postMessage"}}]`);
+  }
+  const list2 = Array.isArray(parsed) ? parsed : [parsed];
+  if (list2.some((item) => !item || typeof item !== "object" || Array.isArray(item))) throw new Error(`Task ${index + 1}: each check must be an object with an id and a state or trace check`);
+  return list2;
+}
+function workspaceDefinition(draft) {
+  return {
+    name: draft.name,
+    agentId: draft.agentId,
+    worldId: draft.worldId,
+    worldRevision: Number(draft.worldRevision),
+    tasks: draft.tasks.map((task, index) => {
+      const checks2 = taskChecks(task, index);
+      return { id: task.id, ...String(task.title ?? "").trim() ? { title: String(task.title).trim() } : {}, prompt: task.prompt, attempts: Number(task.attempts), timeoutMs: Number(task.timeoutMs), ...checks2?.length ? { checks: checks2 } : {} };
+    }),
+    concurrency: Number(draft.concurrency),
+    seed: draft.seed ?? "playground-1"
+  };
+}
+function mountWorldWorkspace(node, options) {
+  const { world, agents, api: api2, go: go2 } = options, base = "/worlds/" + enc(world.id), q = new URLSearchParams(location.search);
+  let tests = [], runs = [], errors = [], selected, agentId = q.get("agent") ?? "", configure = q.has("configure"), showHistory = false, saving = false, disposed = false, error = "", loaded = false;
+  const revisions = /* @__PURE__ */ new Map([[world.revision, world]]);
+  let draft, timer;
+  const launchPending = () => {
+    if (!selected) return false;
+    try {
+      return sessionStorage.getItem("ancha-saved-test-run:" + selected.id) !== null;
+    } catch {
+      return true;
+    }
+  };
+  const key = () => "ancha-world-setup:" + world.id + ":" + (selected?.id ?? "new:" + agentId);
+  const remember = () => {
+    if (draft) try {
+      sessionStorage.setItem(key(), JSON.stringify(draft));
+    } catch {
+    }
+  };
+  function getDraft() {
+    if (draft) return draft;
+    try {
+      const stored = JSON.parse(sessionStorage.getItem(key()) ?? "null");
+      if (stored && Array.isArray(stored.tasks)) draft = stored;
+    } catch {
+    }
+    draft ??= selected ? { ...selected, tasks: structuredClone(selected.tasks) } : { name: world.name, agentId, worldId: world.id, worldRevision: world.revision, concurrency: 1, seed: "playground-1", tasks: [{ id: "task_" + crypto.randomUUID(), title: "", prompt: "", attempts: 1, timeoutMs: 3e5 }] };
+    return draft;
+  }
+  function capture() {
+    const form = node.querySelector(".ww-form");
+    if (!form) return;
+    const d = getDraft(), values = new FormData(form);
+    d.name = String(values.get("name") ?? "");
+    d.agentId = String(values.get("agentId") ?? "");
+    d.concurrency = Number(values.get("concurrency"));
+    d.worldRevision = Number(values.get("worldRevision"));
+    d.tasks = [...form.querySelectorAll(".ww-task-editor")].map((row) => ({ id: row.dataset.taskId, title: row.querySelector("[name=title]").value, prompt: row.querySelector("[name=prompt]").value, attempts: Number(row.querySelector("[name=attempts]").value), timeoutMs: Math.round(Number(row.querySelector("[name=minutes]").value) * 6e4), checksText: row.querySelector("[name=checks]")?.value ?? "" }));
+    remember();
+  }
+  function href2(test, agent = agentId, extra = "") {
+    return base + "?" + new URLSearchParams({ ...agent ? { agent } : {}, ...test ? { test: test.id } : {}, ...extra ? { [extra]: "1" } : {} });
+  }
+  function configuration2() {
+    const d = getDraft(), pinned = Number(d.worldRevision) !== world.revision, startingWorld = revisions.get(Number(d.worldRevision));
+    return `<section class="ww-configure"><div class="ww-section-heading"><h2>Configure</h2><button type="button" data-ww="close">Close</button></div><form class="ww-form"><div class="form-grid"><label>Agent<select name="agentId" required><option value="">Choose agent</option>${agents.map((a) => `<option value="${h(a.id)}" ${d.agentId === a.id ? "selected" : ""}>${h(a.name)}</option>`).join("")}</select></label><label>Configuration name<input name="name" required maxlength="120" value="${h(d.name)}"></label></div><section class="ww-tools"><div class="ww-section-heading"><h3>Tools</h3><a data-nav href="${base}?edit=1">Edit world</a></div><div class="ww-tool-list">${(startingWorld?.configuration.preset.datasets ?? []).map((item) => `<a data-nav href="/tools/${enc(item.tool)}">${toolLogo(item.tool)}${h(toolName(item.tool))}</a>`).join("")}</div>${pinned ? `<label>Starting data<select name="worldRevision"><option value="${h(d.worldRevision)}">Saved revision ${h(d.worldRevision)}</option><option value="${world.revision}">Current revision ${world.revision}</option></select></label>` : `<input type="hidden" name="worldRevision" value="${h(d.worldRevision)}">`}</section><div class="ww-section-heading"><h3>Tasks</h3><button type="button" data-ww="add-task" ${d.tasks.length >= 20 ? "disabled" : ""}>Add task</button></div><div class="ww-task-fields">${d.tasks.map((task, index) => `<article class="ww-task-editor" data-task-id="${h(task.id)}"><div class="ww-section-heading"><h3>Task ${index + 1}</h3>${d.tasks.length > 1 ? `<button type="button" data-ww="remove-task" data-value="${h(task.id)}">Remove</button>` : ""}</div><label>Task name <span class="muted">optional</span><input name="title" maxlength="120" value="${h(task.title ?? "")}" placeholder="Task ${index + 1}"></label><label>Instructions<textarea name="prompt" required maxlength="20000" rows="4">${h(task.prompt)}</textarea></label><div class="ww-task-options"><label>Repeat<input type="number" name="attempts" min="1" max="100" value="${h(task.attempts)}" required><span>times</span></label><label>Time limit <span class="muted">minutes</span><input type="number" name="minutes" min="0.016666666666666666" max="60" step="any" value="${h(task.timeoutMs / 6e4)}" required></label></div><label class="ww-checks-field">Checks <span class="muted">optional \xB7 JSON list, see <a href="/docs#checks" target="_blank" rel="noopener">Docs</a></span><textarea name="checks" rows="${Math.min(12, Math.max(3, checksText(task).split("\n").length))}" spellcheck="false" maxlength="20000" placeholder='[{"id": "posted", "trace": {"called": "POST /api/chat.postMessage"}}]'>${h(checksText(task))}</textarea></label></article>`).join("")}</div><div class="ww-footer"><label>Run at once<input name="concurrency" type="number" min="1" max="10" value="${h(d.concurrency)}" required></label><button type="submit" class="primary" ${saving ? "disabled" : ""}>${saving ? "Saving\u2026" : "Save configuration"}</button></div></form></section>`;
+  }
+  function render() {
+    if (disposed) return;
+    const currentRuns = selected ? setupRuns(runs, selected) : [], run = currentRuns[0], attempts = run?.attempts ?? [], tasks = selected?.tasks ?? [], choices = worldSetups(tests, world.id, agentId);
+    node.innerHTML = `<div class="ww-workspace"><header class="ww-header"><h1>${h(world.name)}</h1><div class="ww-world-actions"><label class="ww-agent-select">Agent<select data-ww-select="agent"><option value="" disabled>Choose agent</option>${agents.map((a) => `<option value="${h(a.id)}" ${a.id === agentId ? "selected" : ""}>${h(a.name)}</option>`).join("")}</select></label><button type="button" data-ww="configure">Configure</button>${selected ? options.runActions?.(selected.id) ?? button("Run tasks", "run-saved-test", selected.id, "primary") : ""}</div></header><p class="ww-error" role="alert" ${error ? "" : "hidden"}>${h(error)}</p>${errors.map((e) => `<p class="ww-error" role="alert">${h(e.message)}</p>`).join("")}${!loaded ? '<p role="status">Loading tasks\u2026</p>' : ""}${choices.length > 1 ? `<label class="ww-setup-select">Configuration<select data-ww-select="test">${!selected ? '<option value="" selected disabled>New configuration</option>' : ""}${choices.map((t) => `<option value="${h(t.id)}" ${t.id === selected?.id ? "selected" : ""}>${h(t.name)}</option>`).join("")}</select></label>` : ""}${loaded && configure ? configuration2() : loaded ? `${summaryWidgets(attempts, "ww-summary-widgets", "article")}<div class="ww-section-heading"><h2>Tasks${tasks.length ? ` <span class="ww-task-count">${tasks.length}</span>` : ""}</h2><div class="actions">${worldHistory(runs, world.id, agentId || void 0).length || selected ? `<button type="button" data-ww="history" aria-expanded="${showHistory}">History</button>` : ""}<button type="button" data-ww="configure">${tasks.length ? "Edit tasks" : "Add tasks"}</button></div></div>${tasks.length ? `<section class="ww-task-table"><div class="ww-table-head"><span>Task</span><span>Success rate</span><span>Attempts</span><span></span></div>${tasks.map((task, index) => {
+      const list2 = attempts.filter((a) => a.taskId === task.id), taskURL = run ? `/playgrounds/${enc(run.id)}?task=${enc(task.id)}${list2[0] ? "&attempt=" + enc(list2[0].id) : ""}` : "";
+      return `<article class="ww-task-row"><div class="ww-task-identity">${taskURL ? `<a data-nav class="ww-task-name" href="${taskURL}"><span class="ww-ordinal">${index + 1}</span><strong>${h(label(task, index))}</strong></a>` : `<strong class="ww-task-name"><span class="ww-ordinal">${index + 1}</span>${h(label(task, index))}</strong>`}<p class="ww-task-prompt" title="${h(task.prompt)}">${h(task.prompt)}</p></div>${(() => {
+        const o = taskOutcome(list2);
+        return `<div class="ww-rate ${o.tone}"><strong>${h(o.value)}</strong><span>${h(o.note)}</span></div>`;
+      })()}<div class="ww-attempt-marks" role="group" aria-label="Attempts for ${h(label(task, index))}">${list2.map((a) => {
+        const o = attemptOutcome(a);
+        return `<a data-nav class="ww-attempt-mark ${h(a.status)} outcome-${o.key}" href="/playgrounds/${enc(run.id)}?task=${enc(task.id)}&attempt=${enc(a.id)}" title="Attempt ${a.number}: ${h(attemptLabel(a.status))} \xB7 ${o.label}" aria-label="Attempt ${a.number}: ${h(attemptLabel(a.status))} \xB7 ${o.label}">${o.key !== "ungraded" ? o.mark : a.status === "error" ? "!" : ["running", "starting"].includes(a.status) ? "\u25CC" : "\u2212"}</a>`;
+      }).join("") || `<span class="muted">Repeat ${h(task.attempts)} times</span>`}</div>${taskURL ? `<a data-nav class="ww-row-inspect" href="${taskURL}" aria-label="Inspect ${h(label(task, index))}">\u2192</a>` : "<span></span>"}</article>`;
+    }).join("")}</section>` : `<div class="quiet-empty">${agents.length ? "Add tasks to this world." : '<a data-nav href="/agents">Connect an agent</a>'}</div>`}${selected && !currentRuns.length && setupHistory(runs, selected).some((item) => item.earlier) ? '<p class="ww-snapshot-notice">This configuration has changed. Earlier attempts are in History.</p>' : ""}${showHistory ? history2() : ""}` : ""}</div>`;
+  }
+  function history2() {
+    const all = worldHistory(runs, world.id, agentId || void 0);
+    return `<section class="ww-history"><div class="ww-section-heading"><h2>History</h2><button type="button" data-ww="new">New configuration</button></div>${all.length ? all.map(({ run, attemptCount, grouped }) => `<a data-nav class="ww-history-row" href="/playgrounds/${enc(run.id)}"><time>${h(new Date(run.createdAt).toLocaleString())}</time><span>${h(attemptLabel(run.status))}</span><span>${attemptCount} attempts</span>${grouped ? "<span>Group</span>" : ""}${run.test?.name || run.groupName ? `<span>${h(run.groupName ?? run.test.name)}</span>` : ""}<span>\u2192</span></a>`).join("") : '<p class="quiet-empty">No attempts yet.</p>'}</section>`;
+  }
+  async function refresh(initial = false) {
+    try {
+      const [definitions, batches] = await Promise.all([api2("/playground-tests"), api2("/playgrounds")]);
+      if (disposed) return;
+      tests = definitions.tests ?? [];
+      runs = batches.runs ?? [];
+      errors = [...definitions.errors ?? [], ...batches.errors ?? []];
+      if (initial) {
+        const requested = q.get("test");
+        if (requested) {
+          selected = worldSetups(tests, world.id).find((test) => test.id === requested);
+          if (!selected || agentId && selected.agentId !== agentId) {
+            selected = void 0;
+            error = "This configuration is unavailable for this world and agent.";
+          } else agentId = selected.agentId;
+        } else if (q.has("new")) {
+          selected = void 0;
+          agentId ||= agents[0]?.id ?? "";
+        } else {
+          selected = worldSetups(tests, world.id, agentId || void 0)[0];
+          const historical = worldHistory(runs, world.id)[0]?.run;
+          agentId ||= selected?.agentId ?? historical?.entries?.find((entry) => entry.world?.id === world.id)?.agentId ?? historical?.agentId ?? agents[0]?.id ?? "";
+        }
+      } else if (selected) selected = tests.find((test) => test.id === selected.id);
+      if (initial && !selected && worldHistory(runs, world.id, agentId || void 0).length) showHistory = true;
+      if (configure) getDraft();
+      options.onContext?.(selected);
+      const pinnedRevision = Number(draft?.worldRevision ?? selected?.worldRevision ?? world.revision);
+      if (!revisions.has(pinnedRevision)) {
+        const savedWorld = await api2(`/worlds/${enc(world.id)}?revision=${pinnedRevision}`);
+        if (disposed) return;
+        revisions.set(pinnedRevision, savedWorld);
+      }
+      loaded = true;
+      if (!configure) render();
+      else if (initial) render();
+      if (runs.some((run) => !terminal(run.status))) timer = setTimeout(() => {
+        void refresh();
+      }, 2e3);
+    } catch (e) {
+      if (!disposed) {
+        error = e instanceof Error ? e.message : String(e);
+        loaded = true;
+        render();
+      }
+    }
+  }
+  node.addEventListener("input", capture);
+  node.addEventListener("change", (event) => {
+    const control = event.target;
+    if (control.dataset.wwSelect === "agent") {
+      capture();
+      const match = worldSetups(tests, world.id, control.value)[0];
+      go2(href2(match, control.value));
+    } else if (control.dataset.wwSelect === "test") go2(href2(tests.find((t) => t.id === control.value)));
+    else {
+      capture();
+      if (control.name === "worldRevision") render();
+    }
+  });
+  node.addEventListener("click", (event) => {
+    const control = event.target.closest("[data-ww]");
+    if (!control) return;
+    event.preventDefault();
+    event.stopPropagation();
+    if (saving) return;
+    capture();
+    error = "";
+    switch (control.dataset.ww) {
+      case "configure":
+        if (launchPending()) {
+          error = "Resolve the pending run before editing this configuration.";
+          break;
+        }
+        configure = true;
+        break;
+      case "close":
+        configure = false;
+        break;
+      case "history":
+        showHistory = !showHistory;
+        break;
+      case "new":
+        go2(href2(void 0, agentId, "new") + "&configure=1");
+        return;
+      case "add-task":
+        if (getDraft().tasks.length < 20) getDraft().tasks.push({ id: "task_" + crypto.randomUUID(), title: "", prompt: "", attempts: 1, timeoutMs: 3e5 });
+        break;
+      case "remove-task":
+        if (getDraft().tasks.length > 1) getDraft().tasks = getDraft().tasks.filter((t) => t.id !== control.dataset.value);
+        break;
+    }
+    remember();
+    render();
+  });
+  node.addEventListener("submit", (event) => {
+    const form = event.target;
+    if (!form.matches(".ww-form")) return;
+    event.preventDefault();
+    event.stopPropagation();
+    if (saving) return;
+    if (launchPending()) {
+      error = "Resolve the pending run before editing this configuration.";
+      render();
+      return;
+    }
+    capture();
+    saving = true;
+    error = "";
+    render();
+    const d = getDraft(), oldKey = key();
+    let body;
+    try {
+      body = workspaceDefinition(d);
+    } catch (e) {
+      error = e instanceof Error ? e.message : String(e);
+      saving = false;
+      render();
+      return;
+    }
+    void api2(selected ? `/playground-tests/${enc(selected.id)}/update` : "/playground-tests", { ...body, ...selected ? { revision: d.revision } : {} }).then((saved) => {
+      sessionStorage.removeItem(oldKey);
+      if (!disposed) go2(href2(saved, saved.agentId));
+    }).catch((e) => {
+      if (!disposed) {
+        error = e instanceof Error ? e.message : String(e);
+        saving = false;
+        render();
+      }
+    });
+  });
+  render();
+  void refresh(true);
+  return () => {
+    disposed = true;
+    clearTimeout(timer);
+  };
 }
 
 // frontend/src/workspace-pages.ts
@@ -1210,7 +1301,8 @@ function mountTestFlow(node, options) {
     if (form) {
       draft.name = form.elements.namedItem("name").value;
       draft.concurrency = Number(form.elements.namedItem("concurrency").value);
-      draft.tasks = [...form.querySelectorAll(".task-editor")].map((row) => ({ id: row.dataset.taskId, prompt: row.querySelector("[name=prompt]").value, attempts: Number(row.querySelector("[name=attempts]").value), timeoutMs: Math.round(Number(row.querySelector("[name=minutes]").value) * 6e4) }));
+      const kept = new Map((draft.tasks ?? []).map((t) => [t.id, t]));
+      draft.tasks = [...form.querySelectorAll(".task-editor")].map((row) => ({ ...kept.get(row.dataset.taskId) ?? {}, id: row.dataset.taskId, prompt: row.querySelector("[name=prompt]").value, attempts: Number(row.querySelector("[name=attempts]").value), timeoutMs: Math.round(Number(row.querySelector("[name=minutes]").value) * 6e4) }));
     }
     remember();
   }
@@ -1474,7 +1566,8 @@ function mountGroupFlow(node, options) {
     const world = options.worlds.find((value) => value.id === draft.worldId);
     draft.worldRevision = world?.revision ?? 0;
     draft.concurrency = Number(values.get("concurrency") ?? 1);
-    draft.tasks = [...form.querySelectorAll(".task-editor")].map((row) => ({ id: row.dataset.taskId, prompt: row.querySelector("[name=prompt]").value, attempts: Number(row.querySelector("[name=attempts]").value), timeoutMs: Math.round(Number(row.querySelector("[name=minutes]").value) * 6e4) }));
+    const kept = new Map((draft.tasks ?? []).map((t) => [t.id, t]));
+    draft.tasks = [...form.querySelectorAll(".task-editor")].map((row) => ({ ...kept.get(row.dataset.taskId) ?? {}, id: row.dataset.taskId, prompt: row.querySelector("[name=prompt]").value, attempts: Number(row.querySelector("[name=attempts]").value), timeoutMs: Math.round(Number(row.querySelector("[name=minutes]").value) * 6e4) }));
     remember();
   }
   function showError(message) {
@@ -1634,10 +1727,10 @@ function comparisonSetup(runId, runs) {
   return `<p class="hint">Choose a completed run with matching scenarios and test data.</p><form id="compare-form"><input type="hidden" name="current" value="${h(runId)}"><label>Baseline<select name="baseline" required>${baselines.map((r) => `<option value="${h(r.runId)}">${h(r.metadata?.label ?? "CLI test")} \xB7 ${h(r.runId)}</option>`).join("")}</select></label><button class="primary">Compare</button></form>`;
 }
 function comparisonResults(result, labels) {
-  const percent = (rate) => `${(rate * 100).toLocaleString("en-US", { maximumFractionDigits: 1 })}%`;
+  const percent2 = (rate) => `${(rate * 100).toLocaleString("en-US", { maximumFractionDigits: 1 })}%`;
   const change = (delta) => delta === 0 ? "No change" : Math.abs(delta * 100) < 0.1 ? `${delta > 0 ? "Increase" : "Decrease"} under 0.1 percentage points` : `${delta > 0 ? "+" : "\u2212"}${Math.abs(delta * 100).toLocaleString("en-US", { maximumFractionDigits: 1 })} percentage points`;
   const attempts = (count) => `${count} ${count === 1 ? "attempt" : "attempts"}`;
-  return `<p>${result.regressed ? "Pass rate decreased in one or more scenarios." : "No decrease in pass rate."}</p><div class="table-wrap comparison-table"><table><thead><tr><th scope="col">Scenario</th><th scope="col">Before</th><th scope="col">After</th><th scope="col">Change</th></tr></thead><tbody>${result.scenes.map((s, index) => `<tr><th scope="row">${h(labels[s.sceneId] ?? `Scenario ${index + 1}`)}</th><td><strong>${percent(s.before)}</strong><span class="comparison-attempts">${attempts(s.beforeTrials)}</span></td><td><strong>${percent(s.after)}</strong><span class="comparison-attempts">${attempts(s.afterTrials)}</span></td><td>${badge(change(s.delta), s.delta > 0 ? "good" : s.delta < 0 ? "bad" : "")}</td></tr>`).join("")}</tbody></table></div><p class="hint">Observed pass rates; no claim of statistical significance.</p>${details("Comparison details", result)}`;
+  return `<p>${result.regressed ? "Pass rate decreased in one or more scenarios." : "No decrease in pass rate."}</p><div class="table-wrap comparison-table"><table><thead><tr><th scope="col">Scenario</th><th scope="col">Before</th><th scope="col">After</th><th scope="col">Change</th></tr></thead><tbody>${result.scenes.map((s, index) => `<tr><th scope="row">${h(labels[s.sceneId] ?? `Scenario ${index + 1}`)}</th><td><strong>${percent2(s.before)}</strong><span class="comparison-attempts">${attempts(s.beforeTrials)}</span></td><td><strong>${percent2(s.after)}</strong><span class="comparison-attempts">${attempts(s.afterTrials)}</span></td><td>${badge(change(s.delta), s.delta > 0 ? "good" : s.delta < 0 ? "bad" : "")}</td></tr>`).join("")}</tbody></table></div><p class="hint">Observed pass rates; no claim of statistical significance.</p>${details("Comparison details", result)}`;
 }
 function dataSelection(item) {
   if (!item) return "";