@recombine-ai/bosun 0.3.6 → 0.3.9

package/build/server/chunks/{_page.svelte-N6xQwQzn.js → _page.svelte-C1YaC_Qg.js}

@@ -1,47 +1,17 @@
-import { h as head, c as attr_style, a as attr_class, b as attr, s as stringify, e as ensure_array_like, g as attributes } from './index2-CHZ7ZFLZ.js';
-import { P as Panel, C as ContextEditor } from './ContextEditor-ySHjHE1w.js';
-import { B as Button } from './Button-D3mqAVvu.js';
-import { M as Modal } from './Modal-NTInCBwo.js';
-import { o as onDestroy } from './Tab.svelte_svelte_type_style_lang-Bmdqxv-M.js';
-import { p as page, g as goto } from './index3-DeV2OWZf.js';
-import { I as Icon$1, h as html } from './Icon-D0R0X6Hr.js';
-import { T as Tabs, a as Tab } from './Tab-DAuTmew0.js';
-import { g as getAgentsContext } from './agents-context-BrlIlkOW.js';
-import { A as AgentSidebar } from './AgentSidebar-C4H-7WxT.js';
-import { I as Input } from './Input-C_H479QL.js';
+import { h as head, b as attr_class, a as attr, s as stringify, e as ensure_array_like, c as attr_style, j as attributes } from './index2-BuOY2QE4.js';
+import { g as getEditWorkspaceContext, P as Panel } from './edit-workspace-context-OTOIRCI0.js';
+import { B as Button } from './Button-Cd4Yzpdy.js';
+import { M as Modal } from './Modal-D4tx0Bdo.js';
+import { C as ContextEditor, P as PopoverModal } from './ContextEditor-CfjBPLpI.js';
+import { I as Input } from './Input-BJW36iKz.js';
+import { o as onDestroy } from './index-server-CefdgakO.js';
+import { I as Icon$1, h as html } from './Icon-BgknHaZs.js';
 import { t as trpc } from './trpc-client-Bucti7HD.js';
 import { X as escape_html } from './context-zH5kvtGY.js';
-import 'js-yaml';
 import './events-M0vvlm3M.js';
-import './state.svelte-DtOGaiYC.js';
-import './utils-KcIDVAAe.js';
-import './index-DxaLoiKh.js';
+import 'js-yaml';
 import '@trpc/client';
 
-function _defineProperty(obj, key, value) {
-  if (key in obj) {
-    Object.defineProperty(obj, key, {
-      value,
-      enumerable: true,
-      configurable: true,
-      writable: true
-    });
-  } else {
-    obj[key] = value;
-  }
-  return obj;
-}
-var FEATURE_FLAG_NAMES = Object.freeze({
-  // This flag exists as a workaround for issue 454 (basically a browser bug) - seems like these rect values take time to update when in grid layout. Setting it to true can cause strange behaviour in the REPL for non-grid zones, see issue 470
-  USE_COMPUTED_STYLE_INSTEAD_OF_BOUNDING_RECT: "USE_COMPUTED_STYLE_INSTEAD_OF_BOUNDING_RECT"
-});
-_defineProperty({}, FEATURE_FLAG_NAMES.USE_COMPUTED_STYLE_INSTEAD_OF_BOUNDING_RECT, false);
-var _ID_TO_INSTRUCTION;
-var INSTRUCTION_IDs$1 = {
-  DND_ZONE_ACTIVE: "dnd-zone-active",
-  DND_ZONE_DRAG_DISABLED: "dnd-zone-drag-disabled"
-};
-_ID_TO_INSTRUCTION = {}, _defineProperty(_ID_TO_INSTRUCTION, INSTRUCTION_IDs$1.DND_ZONE_ACTIVE, "Tab to one the items and press space-bar or enter to start dragging it"), _defineProperty(_ID_TO_INSTRUCTION, INSTRUCTION_IDs$1.DND_ZONE_DRAG_DISABLED, "This is a disabled drag and drop list");
 const matchIconName = /^[a-z0-9]+(-[a-z0-9]+)*$/;
 const stringToIcon = (value, validate, allowSimpleName, provider = "") => {
   const colonSeparated = value.split(":");
@@ -1410,19 +1380,44 @@ function Icon($$renderer, $$props) {
 function MiniEditor($$renderer, $$props) {
   $$renderer.component(($$renderer2) => {
     let editor = null;
+    let model = null;
+    let disposables = [];
+    let editorElement;
     let {
-      value,
-      onChange,
-      onBlur,
-      language = "plaintext",
       height = 120
     } = $$props;
+    function handleWheelCapture(e) {
+      e.stopPropagation();
+    }
     onDestroy(() => {
       try {
-        editor?.dispose();
+        editorElement.removeEventListener("wheel", handleWheelCapture, { capture: true });
+      } catch {
+      }
+      for (const d of disposables) {
+        try {
+          d.dispose();
+        } catch {
+        }
+      }
+      disposables = [];
+      const toDisposeModel = model;
+      model = null;
+      try {
+        if (editor) {
+          try {
+            editor.setModel(null);
+          } catch {
+          }
+          editor.dispose();
+        }
       } finally {
         editor = null;
       }
+      try {
+        toDisposeModel?.dispose();
+      } catch {
+      }
     });
     $$renderer2.push(`<div class="w-full"${attr_style(
       // Update language when prop changes without recreating the editor
@@ -1433,27 +1428,12 @@ function MiniEditor($$renderer, $$props) {
 }
 function _page($$renderer, $$props) {
   $$renderer.component(($$renderer2) => {
-    const { onAgentSelect } = getAgentsContext();
-    let { params } = $$props;
-    let agents = [];
-    const agent = params.agent;
-    let repeats = 1;
-    let concurrency = 1;
-    let defaultContext = {};
-    let defaultContextDraft = {};
-    let showDefaultContextModal = false;
-    let showViewerModal = false;
-    let viewerTitle = "";
-    let viewerContent = null;
-    let viewerLoading = false;
-    let viewerError = null;
-    let showTestContextModal = false;
-    let activeContextTestId = null;
-    let testContextDraft = "";
+    const { state: workspace, actions } = getEditWorkspaceContext();
+    const agent = workspace.agentId;
     function safeClone(v) {
       try {
         return structuredClone(v);
-      } catch (e) {
+      } catch {
         try {
           return JSON.parse(JSON.stringify(v));
         } catch {
@@ -1461,36 +1441,20 @@ function _page($$renderer, $$props) {
         }
       }
     }
-    async function openViewer(title, params2) {
-      viewerTitle = title;
-      viewerContent = null;
-      viewerError = null;
-      viewerLoading = true;
-      showViewerModal = true;
-      try {
-        const text = await trpc.tests.resultEntry.query({
-          agent: params2.agent,
-          testId: params2.id,
-          resultKey: params2.key,
-          timestamp: params2.timestamp,
-          index: params2.run
-        });
-        let formatted = text;
-        try {
-          const json = JSON.parse(text);
-          formatted = JSON.stringify(json, null, 2);
-        } catch {
-        }
-        viewerContent = formatted;
-      } catch (e) {
-        console.error(e);
-        viewerError = "Failed to load content";
-      } finally {
-        viewerLoading = false;
-      }
+    function clamp(n, min, max) {
+      return Math.max(min, Math.min(max, n));
     }
+    let repeats = 1;
+    let concurrency = 1;
+    let defaultContext = {};
+    let defaultContextDraft = {};
+    let showDefaultContextModal = false;
+    let testContextPopoverTestId = null;
+    let testContextDraft = {};
+    let showViewerModal = false;
+    let viewerTitle = "";
+    let analyzeRunning = false;
     const baseAssertionTypes = [
-      // Non-deterministic / model-graded
       { value: "llm-rubric", label: "LLM" },
       { value: "answer-relevance", label: "Answer Relevance (LLM)" },
       {
@@ -1515,7 +1479,6 @@ function _page($$renderer, $$props) {
         value: "conversation-relevance",
         label: "Conversation Relevance"
       },
-      // Deterministic string/content checks
       { value: "equals", label: "Equals" },
       { value: "contains", label: "Contains" },
       { value: "icontains", label: "Contains (case-insensitive)" },
@@ -1527,7 +1490,6 @@ function _page($$renderer, $$props) {
       { value: "regex", label: "Regex" },
       { value: "levenshtein", label: "Levenshtein Distance" },
       { value: "is-refusal", label: "Is Refusal" },
-      // Content type assertions
       { value: "is-html", label: "Is HTML" },
       { value: "contains-html", label: "Contains HTML" },
       { value: "is-json", label: "Is JSON" },
@@ -1536,7 +1498,6 @@ function _page($$renderer, $$props) {
       { value: "contains-xml", label: "Contains XML" },
       { value: "is-sql", label: "Is SQL" },
       { value: "contains-sql", label: "Contains SQL" },
-      // Metrics
       { value: "similar", label: "Semantic Similarity" },
       { value: "rouge-n", label: "ROUGE-N" },
       { value: "bleu", label: "BLEU" },
@@ -1544,14 +1505,12 @@ function _page($$renderer, $$props) {
       { value: "meteor", label: "METEOR" },
       { value: "perplexity", label: "Perplexity" },
       { value: "perplexity-score", label: "Perplexity Score" },
-      // Performance/trace
       { value: "latency", label: "Latency" },
       { value: "cost", label: "Cost" },
       { value: "finish-reason", label: "Finish Reason" },
       { value: "trace-error-spans", label: "Trace Error Spans" },
       { value: "trace-span-count", label: "Trace Span Count" },
       { value: "trace-span-duration", label: "Trace Span Duration" },
-      // Tools/validators
       {
         value: "is-valid-function-call",
         label: "Is Valid Function Call"
@@ -1565,7 +1524,6 @@ function _page($$renderer, $$props) {
         label: "Is Valid OpenAI Tools Call"
       },
       { value: "guardrails", label: "Guardrails" },
-      // Custom hooks
       { value: "javascript", label: "JavaScript (custom)" },
       { value: "python", label: "Python (custom)" },
       { value: "webhook", label: "Webhook (custom)" },
@@ -1577,9 +1535,13 @@ function _page($$renderer, $$props) {
     ];
     let tests = [];
     let running = false;
+    let canceling = false;
+    let cancelRequested = false;
     let runScope = null;
+    let runSingleTestId = null;
     let runDone = 0;
     let runTotal = 0;
+    let runExpectedRuns = 1;
     let pollInFlight = false;
     let pollTimer = null;
     let runToken = 0;
@@ -1597,31 +1559,40 @@ function _page($$renderer, $$props) {
       return p;
     }
     function makeResult(id, rr) {
+      const rawRuns = Array.isArray(rr.results) ? rr.results : [];
+      const successCount = rawRuns.reduce((acc, run2) => acc + (run2?.success ? 1 : 0), 0);
       return {
         success_rate: rr.success_rate,
-        total_runs: Array.isArray(rr.results) ? rr.results.length : 0,
+        successes: successCount,
+        total_runs: rawRuns.length,
         isOld: !!rr.is_old,
-        runs: (Array.isArray(rr.results) ? rr.results : []).map((run2, idx) => {
+        previous: rr.previous && typeof rr.previous === "object" ? {
+          success_count: Number(rr.previous.success_count ?? 0),
+          total_runs: Number(rr.previous.total_runs ?? 0),
+          report_timestamp: String(rr.previous.report_timestamp ?? "")
+        } : void 0,
+        runs: rawRuns.map((run2, idx) => {
           const timestamp = rr.report_timestamp;
-          const params2 = { agent, id, timestamp, run: idx };
+          const params = { agent, id, timestamp, run: idx };
           return {
             index: idx + 1,
             success: !!run2.success,
             text: String(run2.result || ""),
             evaluation: String(run2.evaluation || ""),
             links: {
-              renderedPrompt: { ...params2, key: "rendered_prompt" },
-              receivedPrompt: { ...params2, key: "received_prompt" },
-              conversation: { ...params2, key: "stringified_conversation" },
-              context: { ...params2, key: "received_context" },
-              logs: { ...params2, key: "logs" },
-              traces: { ...params2, key: "traces" }
+              renderedPrompt: { ...params, key: "rendered_prompt" },
+              receivedPrompt: { ...params, key: "received_prompt" },
+              conversation: { ...params, key: "stringified_conversation" },
+              context: { ...params, key: "received_context" },
+              logs: { ...params, key: "logs" },
+              traces: { ...params, key: "traces" }
             }
           };
-        })
+        }).reverse()
       };
     }
     async function refresh() {
+      if (!agent) return;
       try {
         const [{ rows }, config, results] = await Promise.all([
           trpc.tests.list.query({ agent }),
@@ -1638,18 +1609,21 @@ function _page($$renderer, $$props) {
             id,
             pinned: String(r.pinned).toUpperCase() === "TRUE",
             scenario: r.scenario || "",
+            step: String(r.step ?? ""),
             author: r.author || "",
             assertion_type: r.assertion_type || "llm-rubric",
             assertion_value: r.assertion_value || "",
             conversation: r.conversation || "",
             expected_output: r.expected_output || "",
+            conversation_ui: buildConversationEditorValue(r.conversation || "", r.expected_output || ""),
             timeout: Number(r.timeout ?? 0),
             similarity_threshold: Number(r.similarity_threshold ?? 60),
             context: JSON.stringify(r.context),
             result: rr ? makeResult(id, rr) : void 0
           };
         });
-      } catch {
+      } catch (e) {
+        console.error(e);
         tests = [];
         repeats = 1;
         concurrency = 1;
@@ -1670,10 +1644,12 @@ function _page($$renderer, $$props) {
         });
         tests = nextTests;
         const scope = runScope;
+        const expectedRuns = Math.max(1, Number(runExpectedRuns) || 1);
         const done = nextTests.reduce(
           (acc, t) => {
+            if (scope === "single" && t.id !== runSingleTestId) return acc;
             if (scope === "pinned" && !t.pinned) return acc;
-            return t.result && !t.result.isOld ? acc + 1 : acc;
+            return t.result && !t.result.isOld && t.result.total_runs >= expectedRuns ? acc + 1 : acc;
           },
           0
         );
@@ -1686,9 +1662,12 @@ function _page($$renderer, $$props) {
     }
     async function run(scope) {
       running = true;
+      cancelRequested = false;
       runScope = scope;
+      runSingleTestId = null;
       runTotal = scope === "all" ? tests.length : tests.filter((t) => t.pinned).length;
       runDone = 0;
+      runExpectedRuns = Math.max(1, Number(repeats) || 1);
       runToken += 1;
       const activeToken = runToken;
       await Promise.all([...pendingSaves]);
@@ -1708,10 +1687,50 @@ function _page($$renderer, $$props) {
         }
       }
     }
+    async function runSingle(testId) {
+      if (!agent) return;
+      running = true;
+      cancelRequested = false;
+      runScope = "single";
+      runSingleTestId = testId;
+      runTotal = 1;
+      runDone = 0;
+      runExpectedRuns = Math.max(1, Number(repeats) || 1);
+      runToken += 1;
+      const activeToken = runToken;
+      await Promise.all([...pendingSaves]);
+      if (pollTimer) clearInterval(pollTimer);
+      pollTimer = setInterval(() => pollLatestResults(activeToken), 800);
+      void pollLatestResults(activeToken);
+      try {
+        await trpc.tests.runTests.mutate({ agent, testId });
+      } finally {
+        if (pollTimer) clearInterval(pollTimer);
+        pollTimer = null;
+        try {
+          await refresh();
+        } finally {
+          running = false;
+          runScope = null;
+          runSingleTestId = null;
+        }
+      }
+    }
+    async function cancelCurrentRun() {
+      if (!running || canceling || cancelRequested) return;
+      canceling = true;
+      try {
+        const ok = await trpc.tests.cancelRun.mutate({ agent });
+        if (ok) cancelRequested = true;
+      } catch (e) {
+        console.error(e);
+      } finally {
+        canceling = false;
+      }
+    }
     const pinnedCount = tests.filter((t) => t.pinned).length;
     const hasPinned = pinnedCount > 0;
-    const allPinned = tests.length > 0 && pinnedCount === tests.length;
-    const somePinned = pinnedCount > 0 && !allPinned;
+    tests.length > 0 && pinnedCount === tests.length;
     async function addTest() {
       const res = await trpc.tests.addTest.mutate({ agent, seed: {}, mergeContexts: true });
       tests = [parseTest(res), ...tests];
@@ -1719,15 +1738,14 @@ function _page($$renderer, $$props) {
     function parseTest(res) {
       return {
         ...res,
+        step: String(res.step ?? ""),
         pinned: res.pinned === "TRUE",
         timeout: Number(res.timeout),
         similarity_threshold: Number(res.similarity_threshold),
-        context: JSON.stringify(res.context)
+        context: JSON.stringify(res.context),
+        conversation_ui: buildConversationEditorValue(res.conversation || "", res.expected_output || "")
       };
     }
-    function clamp(n, min, max) {
-      return Math.max(min, Math.min(max, n));
-    }
     function formatEvaluation(ev) {
       if (!ev) return "";
       try {
@@ -1743,58 +1761,138 @@ function _page($$renderer, $$props) {
         return {};
       }
     }
-    const saveTimers = /* @__PURE__ */ new Map();
-    function saveRowDebounced(id, patch, delay = 600) {
-      const key = `${id}:context`;
-      const prev = saveTimers.get(key);
-      if (prev) clearTimeout(prev);
-      const t = setTimeout(
-        () => {
-          trackSave(saveRow(id, patch));
-          saveTimers.delete(key);
-        },
-        delay
-      );
-      saveTimers.set(key, t);
+    const roleMatch = (line) => line.match(/^\s*(agent|assistant|user|system)\s*:\s*(.*)$/i);
+    function formatExpectedAsAgentBlock(expected) {
+      const trimmed = String(expected ?? "").trimEnd();
+      if (!trimmed) return "";
+      if (/^\s*(agent|assistant)\s*:/i.test(trimmed)) return trimmed;
+      return `Agent: ${trimmed}`;
+    }
+    function buildConversationEditorValue(conversation, expected) {
+      const conv = String(conversation ?? "").trimEnd();
+      const expectedBlock = formatExpectedAsAgentBlock(expected);
+      if (!expectedBlock) return conv;
+      if (!conv) return expectedBlock;
+      return `${conv}
+${expectedBlock}`;
+    }
+    function parseExpectedFromConversationEditor(text) {
+      const raw = String(text ?? "");
+      const lines = raw.split(/\r?\n/);
+      const blocks = [];
+      let current = null;
+      for (let i = 0; i < lines.length; i++) {
+        const m2 = roleMatch(lines[i]);
+        if (!m2) continue;
+        if (current) {
+          blocks.push({ ...current, endLine: i });
+        }
+        const roleRaw = String(m2[1] ?? "").toLowerCase();
+        const role = roleRaw === "assistant" ? "agent" : roleRaw;
+        current = { role, startLine: i + 1 };
+      }
+      if (current) blocks.push({ ...current, endLine: lines.length });
+      const expectedBlock = blocks.at(-1);
+      const promptBlock = blocks.length >= 2 ? blocks.at(-2) : null;
+      const canExtractExpected = expectedBlock?.role === "agent";
+      const valid = canExtractExpected && promptBlock?.role === "user";
+      if (!canExtractExpected || !expectedBlock) {
+        return { valid: false, expected: "", conversation: raw.trimEnd() };
+      }
+      const expectedLines = lines.slice(expectedBlock.startLine - 1, expectedBlock.endLine);
+      const first = expectedLines[0] ?? "";
+      const m = roleMatch(first);
+      const firstContent = m ? m[2] ?? "" : "";
+      const expected = [firstContent, ...expectedLines.slice(1)].join("\n").trimEnd();
+      const conversation = lines.slice(0, expectedBlock.startLine - 1).join("\n").trimEnd();
+      return {
+        valid,
+        expected,
+        conversation,
+        highlight: {
+          startLine: expectedBlock.startLine,
+          endLine: expectedBlock.endLine
+        }
+      };
+    }
+    function erf(x) {
+      const sign = x < 0 ? -1 : 1;
+      const ax = Math.abs(x);
+      const a1 = 0.254829592;
+      const a2 = -0.284496736;
+      const a3 = 1.421413741;
+      const a4 = -1.453152027;
+      const a5 = 1.061405429;
+      const p = 0.3275911;
+      const t = 1 / (1 + p * ax);
+      const y = 1 - ((((a5 * t + a4) * t + a3) * t + a2) * t + a1) * t * Math.exp(-ax * ax);
+      return sign * y;
+    }
+    function normalCdf(x) {
+      return 0.5 * (1 + erf(x / Math.SQRT2));
+    }
+    function twoProportionZTestPValue(x1, n1, x2, n2) {
+      if (!Number.isFinite(x1) || !Number.isFinite(n1) || !Number.isFinite(x2) || !Number.isFinite(n2)) return null;
+      if (n1 <= 0 || n2 <= 0) return null;
+      const p1 = x1 / n1;
+      const p2 = x2 / n2;
+      const pooled = (x1 + x2) / (n1 + n2);
+      const denom = Math.sqrt(pooled * (1 - pooled) * (1 / n1 + 1 / n2));
+      if (!Number.isFinite(denom) || denom <= 0) return null;
+      const z = (p2 - p1) / denom;
+      const p = 2 * (1 - normalCdf(Math.abs(z)));
+      return Math.max(0, Math.min(1, p));
+    }
+    function getExpectedTrend(result, expectedRuns) {
+      if (!result || result.isOld) return null;
+      const doneRuns = result.total_runs;
+      if (doneRuns <= 0) return null;
+      const prev = result.previous;
+      const prevRate = prev && prev.total_runs > 0 ? prev.success_count / prev.total_runs : null;
+      const currRate = result.successes / doneRuns;
+      const direction = prevRate === null ? "higher" : currRate >= prevRate ? "higher" : "lower";
+      const significant = (() => {
+        if (doneRuns <= 5 || !prev || prev.total_runs <= 0) return false;
+        const pValue = twoProportionZTestPValue(prev.success_count, prev.total_runs, result.successes, doneRuns);
+        return pValue !== null && pValue < 0.05;
+      })();
+      const showExpected = Number.isFinite(expectedRuns) && expectedRuns > 0 && doneRuns < expectedRuns;
+      if (!showExpected && !significant) return null;
+      const expectedSuccesses = showExpected ? Math.round(result.successes / doneRuns * expectedRuns) : void 0;
+      return {
+        showExpected,
+        expectedSuccesses,
+        expectedRuns,
+        significant,
+        direction
+      };
     }
     async function saveRow(id, patch) {
       await trpc.tests.patchTest.mutate({ agent, id, patch });
     }
-    async function saveConfigPatch(patch) {
-      trpc.tests.updateConfig.mutate({ agent: params.agent, cfg: patch });
+    function openTestContext(row) {
+      if (!agent) return;
+      testContextPopoverTestId = row.id;
+      testContextDraft = safeClone(parseJson(row.context));
     }
-    const dec = (s) => {
-      try {
-        return s ? decodeURIComponent(s) : "";
-      } catch {
-        return s || "";
-      }
-    };
-    const currentAgentNorm = (() => {
-      return dec(agent).trim().replace(/^\/+|\/+$/g, "");
-    })();
-    const activeTopTab = (() => page.url.pathname.split("/")[1])();
-    function handleTabChange({ value }) {
-      if (value === activeTopTab) return;
-      page.url.searchParams.get("step");
-      if (value === "roleplay") {
-        goto();
-        return;
-      }
-      if (value === "tests") {
-        goto();
+    async function closeTestContextPopover() {
+      const id = testContextPopoverTestId;
+      if (!id || !agent) {
+        testContextPopoverTestId = null;
         return;
       }
-      if (value === "traces") {
-        goto();
-        return;
+      const draft = safeClone(testContextDraft);
+      testContextPopoverTestId = null;
+      tests = tests.map((t) => t.id === id ? { ...t, context: JSON.stringify(draft) } : t);
+      try {
+        await trackSave(saveRow(id, { context: draft }));
+      } catch (e) {
+        console.error(e);
+        await refresh();
       }
-      goto();
     }
-    function handleAgentClick(agentId) {
-      onAgentSelect(agentId);
-      const [proj, ag] = (agentId || "").split("/");
-      if (proj && ag) goto();
+    async function saveConfigPatch(patch) {
+      await trpc.tests.updateConfig.mutate({ agent, cfg: patch });
     }
     onDestroy(() => {
       if (pollTimer) clearInterval(pollTimer);
@@ -1803,149 +1901,8 @@ function _page($$renderer, $$props) {
     let $$inner_renderer;
     function $$render_inner($$renderer3) {
       head($$renderer3, ($$renderer4) => {
-        $$renderer4.title(($$renderer5) => {
-          $$renderer5.push(`<title>Recombine: Tests</title>`);
-        });
         $$renderer4.push(`<link rel="preload" href="/ui/arrow_tests.svg" as="image"/>`);
       });
-      $$renderer3.push(`<div class="flex h-screen">`);
-      AgentSidebar($$renderer3, {
-        className: "flex-none",
-        agents,
-        currentAgent: currentAgentNorm,
-        onAgentClick: handleAgentClick
-      });
-      $$renderer3.push(`<!----> <div class="flex h-full min-w-0 flex-1 flex-col bg-rc-gray-980 pb-3 pr-4 pt-2"><nav class="mb-3 flex h-14 flex-none items-center justify-between bg-rc-gray-980"><div class="flex items-center">`);
-      Tabs($$renderer3, {
-        value: activeTopTab,
-        ariaLabel: "Top tabs",
-        onChange: handleTabChange,
-        children: ($$renderer4) => {
-          Tab($$renderer4, {
-            value: "roleplay",
-            children: ($$renderer5) => {
-              $$renderer5.push(`<!---->Roleplay`);
-            }
-          });
-          $$renderer4.push(`<!----> `);
-          Tab($$renderer4, {
-            value: "tests",
-            children: ($$renderer5) => {
-              $$renderer5.push(`<!---->Tests`);
-            }
-          });
-          $$renderer4.push(`<!----> `);
-          Tab($$renderer4, {
-            value: "traces",
-            children: ($$renderer5) => {
-              $$renderer5.push(`<!---->Traces`);
-            }
-          });
-          $$renderer4.push(`<!---->`);
-        },
-        $$slots: { default: true }
-      });
-      $$renderer3.push(`<!----></div> `);
-      Panel($$renderer3, {
-        class: "!rounded-none p-3 pr-0",
-        children: ($$renderer4) => {
-          $$renderer4.push(`<div class="flex items-center gap-16"><div class="flex items-center">`);
-          Button($$renderer4, {
-            variant: "ghost",
-            class: "text-sm font-semibold text-rc-gray-400 underline decoration-dashed underline-offset-4 hover:text-rc-gray-200",
-            onclick: () => {
-              defaultContextDraft = safeClone(defaultContext);
-              showDefaultContextModal = true;
-            },
-            children: ($$renderer5) => {
-              $$renderer5.push(`<!---->Default context`);
-            },
-            $$slots: { default: true }
-          });
-          $$renderer4.push(`<!----> <div class="relative inline-flex items-center">`);
-          Button($$renderer4, {
-            variant: "secondary",
-            onclick: addTest,
-            children: ($$renderer5) => {
-              $$renderer5.push(`<!---->New test`);
-            },
-            $$slots: { default: true }
-          });
-          $$renderer4.push(`<!----> `);
-          if (hasFetchedTests && tests.length === 0) {
-            $$renderer4.push("<!--[-->");
-            $$renderer4.push(`<div class="pointer-events-none absolute -left-28 top-full mt-4 inline-flex -translate-x-1/2 flex-col items-center"><div class="relative w-50vw max-w-2xl"><img src="/ui/arrow_tests.svg" alt="Arrow" class="block h-auto w-full object-contain"/> <span class="absolute inset-0 mt-48 flex items-center justify-center text-center text-4xl font-extrabold text-rc-gray-200">No tests yet.<br/>Add the first one</span></div></div>`);
-          } else {
-            $$renderer4.push("<!--[!-->");
-          }
-          $$renderer4.push(`<!--]--></div></div> <div class="flex flex-wrap items-center gap-3">`);
-          if (running) {
-            $$renderer4.push("<!--[-->");
-            $$renderer4.push(`<div class="mr-3 flex items-center gap-3"><span class="animate-pulse rounded bg-rc-green-200 px-2 py-1 text-xs text-white">Running… ${escape_html(runDone)}/${escape_html(runTotal)}</span></div>`);
-          } else {
-            $$renderer4.push("<!--[!-->");
-          }
-          $$renderer4.push(`<!--]--> <div class="flex items-center gap-2"><label class="font-semibold text-rc-gray-300">Repeats</label> `);
-          Input($$renderer4, {
-            type: "number",
-            min: "0",
-            max: "500",
-            class: "w-16",
-            onchange: () => {
-              repeats = clamp(+repeats || 0, 0, 500);
-              trackSave(saveConfigPatch({ repeats }));
-            },
-            get value() {
-              return repeats;
-            },
-            set value($$value) {
-              repeats = $$value;
-              $$settled = false;
-            }
-          });
-          $$renderer4.push(`<!----></div> <div class="flex items-center gap-2"><label class="font-semibold text-rc-gray-300">Concurrency</label> `);
-          Input($$renderer4, {
-            type: "number",
-            min: "0",
-            max: "20",
-            class: "w-16",
-            onchange: () => {
-              concurrency = clamp(+concurrency || 0, 0, 20);
-              trackSave(saveConfigPatch({ concurrency }));
-            },
-            get value() {
-              return concurrency;
-            },
-            set value($$value) {
-              concurrency = $$value;
-              $$settled = false;
-            }
-          });
-          $$renderer4.push(`<!----></div> `);
-          Button($$renderer4, {
-            variant: "secondary",
-            class: "disabled:cursor-not-allowed disabled:opacity-50",
-            disabled: !hasPinned || running || isSaving,
-            onclick: () => run("pinned"),
-            children: ($$renderer5) => {
-              $$renderer5.push(`<!---->Run pinned`);
-            },
-            $$slots: { default: true }
-          });
-          $$renderer4.push(`<!----> `);
-          Button($$renderer4, {
-            class: "disabled:cursor-not-allowed disabled:opacity-50",
-            disabled: running || isSaving || tests.length === 0,
-            onclick: () => run("all"),
-            children: ($$renderer5) => {
-              $$renderer5.push(`<!---->Run all`);
-            },
-            $$slots: { default: true }
-          });
-          $$renderer4.push(`<!----></div></div>`);
-        }
-      });
-      $$renderer3.push(`<!----></nav>  `);
       {
         let children = function($$renderer4) {
           ContextEditor($$renderer4, {
@@ -1960,9 +1917,7 @@ function _page($$renderer, $$props) {
           Button($$renderer4, {
             variant: "outlined",
             size: "small",
-            onclick: () => {
-              showDefaultContextModal = false;
-            },
+            onclick: () => showDefaultContextModal = false,
             children: ($$renderer5) => {
               $$renderer5.push(`<!---->Cancel`);
             },
@@ -2000,51 +1955,34 @@ function _page($$renderer, $$props) {
       $$renderer3.push(`<!----> `);
       {
         let children = function($$renderer4) {
-          MiniEditor($$renderer4, {
-            value: testContextDraft,
-            onChange: (v) => testContextDraft = v,
-            language: "json",
-            height: 360
-          });
+          {
+            $$renderer4.push("<!--[!-->");
+            {
+              $$renderer4.push("<!--[!-->");
+              $$renderer4.push(`<pre class="whitespace-pre-wrap font-mono text-gray-200">${escape_html("—")}</pre>`);
+            }
+            $$renderer4.push(`<!--]-->`);
+          }
+          $$renderer4.push(`<!--]-->`);
         }, footer = function($$renderer4) {
           Button($$renderer4, {
             variant: "outlined",
             size: "small",
-            onclick: () => {
-              showTestContextModal = false;
-              activeContextTestId = null;
-            },
-            children: ($$renderer5) => {
-              $$renderer5.push(`<!---->Cancel`);
-            },
-            $$slots: { default: true }
-          });
-          $$renderer4.push(`<!----> `);
-          Button($$renderer4, {
-            size: "small",
-            onclick: async () => {
-              const row = tests.find((r) => r.id === activeContextTestId);
-              if (row) {
-                row.context = testContextDraft;
-                await trackSave(saveRow(row.id, { context: JSON.parse(row.context) }));
-              }
-              showTestContextModal = false;
-              activeContextTestId = null;
-            },
+            onclick: () => showViewerModal = false,
             children: ($$renderer5) => {
-              $$renderer5.push(`<!---->Save`);
+              $$renderer5.push(`<!---->Close`);
             },
             $$slots: { default: true }
           });
-          $$renderer4.push(`<!---->`);
         };
         Modal($$renderer3, {
-          title: "Edit Test Context",
+          title: viewerTitle,
+          class: "!w-[80vw] !max-w-5xl",
           get open() {
-            return showTestContextModal;
+            return showViewerModal;
           },
           set open($$value) {
-            showTestContextModal = $$value;
+            showViewerModal = $$value;
             $$settled = false;
           },
           children,
@@ -2052,69 +1990,152 @@ function _page($$renderer, $$props) {
           $$slots: { default: true, footer: true }
         });
       }
+      $$renderer3.push(`<!----> <div class="flex h-full min-h-0 flex-col"><div class="flex flex-none items-center justify-between gap-3 border-b border-rc-gray-700 p-2"><div class="flex items-center gap-3">`);
+      Button($$renderer3, {
+        variant: "secondary",
+        disabled: running || isSaving || !agent,
+        onclick: () => {
+          defaultContextDraft = safeClone(defaultContext);
+          showDefaultContextModal = true;
+        },
+        children: ($$renderer4) => {
+          $$renderer4.push(`<!---->Edit default context`);
+        },
+        $$slots: { default: true }
+      });
+      $$renderer3.push(`<!----> <div class="relative inline-flex items-center">`);
+      Button($$renderer3, {
+        variant: "secondary",
+        onclick: addTest,
+        disabled: isSaving || running || !agent,
+        children: ($$renderer4) => {
+          $$renderer4.push(`<!---->New test`);
+        },
+        $$slots: { default: true }
+      });
       $$renderer3.push(`<!----> `);
-      {
-        let children = function($$renderer4) {
-          if (viewerLoading) {
+      if (hasFetchedTests && tests.length === 0) {
+        $$renderer3.push("<!--[-->");
+        $$renderer3.push(`<div class="pointer-events-none absolute -left-28 top-full mt-4 inline-flex -translate-x-1/2 flex-col items-center"><div class="relative w-50vw max-w-2xl"><img src="/ui/arrow_tests.svg" alt="Arrow" class="block h-auto w-full object-contain"/> <span class="absolute inset-0 mt-48 flex items-center justify-center text-center text-4xl font-extrabold text-rc-gray-200">No tests yet.<br/>Add the first one</span></div></div>`);
+      } else {
+        $$renderer3.push("<!--[!-->");
+      }
+      $$renderer3.push(`<!--]--></div></div> <div class="flex flex-wrap items-center gap-3">`);
+      if (running) {
+        $$renderer3.push("<!--[-->");
+        $$renderer3.push(`<div class="mr-3 flex items-center gap-3"><span class="animate-pulse rounded bg-rc-green-200 px-3 py-1 text-white">Running… ${escape_html(runDone)}/${escape_html(runTotal)}</span> `);
+        Button($$renderer3, {
+          variant: "secondary",
+          disabled: canceling || cancelRequested,
+          onclick: cancelCurrentRun,
+          children: ($$renderer4) => {
+            $$renderer4.push(`<!---->${escape_html(cancelRequested ? "Cancel requested…" : canceling ? "Canceling…" : "Cancel")}`);
+          },
+          $$slots: { default: true }
+        });
+        $$renderer3.push(`<!----></div>`);
+      } else {
+        $$renderer3.push("<!--[!-->");
+      }
+      $$renderer3.push(`<!--]--> <div class="flex items-center gap-2"><label class="text-rc-gray-300">Repeats</label> `);
+      Input($$renderer3, {
+        type: "number",
+        min: "0",
+        max: "500",
+        class: "w-16",
+        onchange: () => {
+          repeats = clamp(+repeats || 0, 0, 500);
+          trackSave(saveConfigPatch({ repeats }));
+        },
+        get value() {
+          return repeats;
+        },
+        set value($$value) {
+          repeats = $$value;
+          $$settled = false;
+        }
+      });
+      $$renderer3.push(`<!----></div> <div class="flex items-center gap-2"><label class="text-rc-gray-300">Concurrency</label> `);
+      Input($$renderer3, {
+        type: "number",
+        min: "0",
+        max: "20",
+        class: "w-16",
+        onchange: () => {
+          concurrency = clamp(+concurrency || 0, 0, 20);
+          trackSave(saveConfigPatch({ concurrency }));
+        },
+        get value() {
+          return concurrency;
+        },
+        set value($$value) {
+          concurrency = $$value;
+          $$settled = false;
+        }
+      });
+      $$renderer3.push(`<!----></div> `);
+      Button($$renderer3, {
+        class: "disabled:cursor-not-allowed disabled:opacity-50",
+        disabled: running || isSaving || tests.length === 0 || hasPinned && pinnedCount === 0,
+        onclick: () => run(hasPinned ? "pinned" : "all"),
+        children: ($$renderer4) => {
+          $$renderer4.push(`<!---->▶ ${escape_html(hasPinned ? "Run pinned" : "Run all")}`);
+        },
+        $$slots: { default: true }
+      });
+      $$renderer3.push(`<!----></div></div> <div class="min-h-0 flex-1 overflow-hidden p-2">`);
+      Panel($$renderer3, {
+        class: "relative flex h-full min-w-0 flex-col overflow-hidden !rounded-none p-0",
+        children: ($$renderer4) => {
+          $$renderer4.push(`<div class="rc-scroll min-h-0 flex-1 overflow-auto"><div class="grid grid-cols-[2.25rem,1fr] items-center gap-2 border-b border-rc-gray-700 p-2"><button${attr_class(
+            `icon-btn justify-self-center ${stringify(hasPinned ? "text-rc-green-350 hover:text-rc-green-250" : "text-gray-300 hover:text-gray-100")}`,
+            "svelte-rdft55"
+          )} aria-label="Pin/unpin all" title="Pin/unpin all"${attr("disabled", running || isSaving || tests.length === 0, true)}>`);
+          Icon($$renderer4, {
+            class: "block",
+            icon: hasPinned ? "mdi:pin" : "mdi:pin-outline",
+            width: 18,
+            height: 18
+          });
+          $$renderer4.push(`<!----></button> <span class="text-left text-rc-gray-300">Pinned: ${escape_html(pinnedCount)}/${escape_html(tests.length)}</span></div> `);
+          if (!hasFetchedTests) {
             $$renderer4.push("<!--[-->");
-            $$renderer4.push(`<div class="text-sm text-gray-300">Loading…</div>`);
+            $$renderer4.push(`<div class="p-3 text-gray-400">Loading tests…</div>`);
           } else {
             $$renderer4.push("<!--[!-->");
-            if (viewerError) {
+            if (tests.length === 0) {
               $$renderer4.push("<!--[-->");
-              $$renderer4.push(`<div class="text-sm text-red-400">${escape_html(viewerError)}</div>`);
+              $$renderer4.push(`<div class="p-3 text-gray-400">—</div>`);
             } else {
               $$renderer4.push("<!--[!-->");
-              $$renderer4.push(`<pre class="whitespace-pre-wrap font-mono text-sm text-gray-200">${escape_html(viewerContent)}</pre>`);
-            }
-            $$renderer4.push(`<!--]-->`);
-          }
-          $$renderer4.push(`<!--]-->`);
-        };
-        Modal($$renderer3, {
-          title: viewerTitle,
-          class: "w-fit min-w-[560px]",
-          get open() {
-            return showViewerModal;
-          },
-          set open($$value) {
-            showViewerModal = $$value;
-            $$settled = false;
-          },
-          children,
-          $$slots: { default: true }
-        });
-      }
-      $$renderer3.push(`<!----> <div class="rc-scroll h-[calc(100vh-3.5rem)] overflow-auto p-0">`);
-      if (!hasFetchedTests) {
-        $$renderer3.push("<!--[-->");
-        $$renderer3.push(`<div class="flex h-full items-center justify-center text-sm text-gray-400">Loading tests…</div>`);
-      } else {
-        $$renderer3.push("<!--[!-->");
-        if (tests.length > 0) {
-          $$renderer3.push("<!--[-->");
-          Panel($$renderer3, {
-            class: "overflow-hidden !rounded-none p-0",
-            children: ($$renderer4) => {
-              $$renderer4.push(`<div class="overflow-auto"><table class="min-w-full text-sm"><thead><tr class="align-top"><th class="w-20 p-0 text-center"><div class="flex items-center justify-end pr-2"><span class="font-normal">Pin all</span> <button${attr_class(`inline-flex h-8 w-8 items-center justify-center rounded disabled:cursor-not-allowed disabled:opacity-50 ${stringify(allPinned || somePinned ? "text-rc-green-400 hover:text-rc-green-300" : "text-gray-400 hover:text-gray-200")}`)} aria-label="Pin/unpin all" title="Pin/unpin all"${attr("disabled", running || isSaving || tests.length === 0, true)}>`);
-              Icon($$renderer4, {
-                icon: allPinned ? "mdi:pin" : "mdi:pin-outline",
-                width: 18,
-                height: 18,
-                class: "relative top-[1px]"
-              });
-              $$renderer4.push(`<!----></button></div></th><th class="p-0"><div class="p-2"><div class="grid grid-cols-1 gap-3 md:grid-cols-12"><div class="flex items-center justify-between md:col-span-4"><span class="text-left font-semibold text-rc-gray-300">Pinned: ${escape_html(pinnedCount)}/${escape_html(tests.length)}</span> <span class="mr-3 text-right font-semibold uppercase tracking-wide text-gray-300">Test</span></div> <div class="ml-3 text-left font-semibold uppercase tracking-wide text-gray-300 md:col-span-8">Result</div></div></div></th></tr></thead><tbody><!--[-->`);
+              $$renderer4.push(`<div class="flex flex-col space-y-5"><!--[-->`);
               const each_array = ensure_array_like(tests);
               for (let i = 0, $$length = each_array.length; i < $$length; i++) {
                 let t = each_array[i];
-                $$renderer4.push(`<tr class="align-top"><td class="w-20 p-0 text-center align-top"><div class="p-2"><div class="flex items-center justify-center gap-1"><span class="inline-flex h-8 w-8 cursor-grab select-none items-center justify-center rounded text-xl leading-none text-gray-400 hover:text-gray-200 active:cursor-grabbing" title="Drag row" aria-label="Drag row">⠿</span> <button${attr_class(`inline-flex h-8 w-8 items-center justify-center rounded disabled:cursor-not-allowed disabled:opacity-50 ${stringify(t.pinned ? "text-rc-green-400 hover:text-rc-green-300" : "text-gray-400 hover:text-gray-200")}`)}${attr("aria-label", t.pinned ? "Unpin" : "Pin")}${attr("title", t.pinned ? "Unpin" : "Pin")}${attr("disabled", running || isSaving, true)}>`);
+                const expectedInfo = parseExpectedFromConversationEditor(t.conversation_ui);
+                $$renderer4.push(`<div class="border-b border-rc-gray-700 p-2"${attr("data-id", t.id)}><div class="grid grid-cols-[2.25rem,1fr] gap-2"><div class="flex flex-col items-center gap-1"><button${attr_class(
+                  `icon-btn ${stringify(t.pinned ? "text-rc-green-350 hover:text-rc-green-250" : "text-gray-300 hover:text-gray-100")}`,
+                  "svelte-rdft55"
+                )}${attr("aria-label", t.pinned ? "Unpin" : "Pin")}${attr("title", t.pinned ? "Unpin" : "Pin")}${attr("disabled", running || isSaving, true)}>`);
                 Icon($$renderer4, {
+                  class: "block",
                   icon: t.pinned ? "mdi:pin" : "mdi:pin-outline",
                   width: 18,
-                  height: 18,
-                  class: "relative top-[1px]"
+                  height: 18
                 });
-                $$renderer4.push(`<!----></button></div></div></td><td class="p-0"><div class="border-b border-rc-gray-700 p-2"><div class="mt-2 grid grid-cols-1 gap-3 md:grid-cols-12"><div class="self-start md:col-span-4"><div class="left-details-bg flex flex-col gap-2 rounded border border-rc-gray-700 p-2 svelte-f5z0gs"><div class="mb-1 flex justify-between"><div><input class="card-input-title svelte-f5z0gs" type="text" placeholder="Scenario"${attr("value", t.scenario)}/> <input class="card-input-subtle mt-1 svelte-f5z0gs" type="text" placeholder="Author"${attr("value", t.author)}/></div> <button title="Delete" class="m-2 inline-flex shrink-0 text-xl text-rc-gray-200 hover:text-rc-gray-100">🗑</button></div> <div class="flex items-center gap-2">`);
+                $$renderer4.push(`<!----></button> <button title="Duplicate" class="icon-btn text-gray-300 hover:text-gray-100 disabled:cursor-not-allowed disabled:opacity-50 svelte-rdft55"${attr("disabled", running || isSaving, true)}>`);
+                Icon($$renderer4, { icon: "mdi:content-copy", width: 18, height: 18 });
+                $$renderer4.push(`<!----></button> <button title="Delete" class="icon-btn text-gray-300 hover:text-gray-100 disabled:cursor-not-allowed disabled:opacity-50 svelte-rdft55"${attr("disabled", running || isSaving, true)}>`);
+                Icon($$renderer4, { icon: "mdi:trash-can-outline", width: 18, height: 18 });
+                $$renderer4.push(`<!----></button></div> <div class="min-w-0"><div${attr_class(`left-details-bg flex flex-col rounded border ${expectedInfo.valid ? "border-rc-gray-700" : "border-rc-red-600"}`, "svelte-rdft55")}><div class="left-details-bg sticky top-0 z-10 rounded-t border-b border-rc-gray-700 px-2 pb-1 pt-2 svelte-rdft55"><input class="card-input-title svelte-rdft55" type="text" placeholder="Scenario"${attr("value", t.scenario)}/></div> <div class="flex flex-col gap-2 p-2">`);
+                if (!expectedInfo.valid) {
+                  $$renderer4.push("<!--[-->");
+                  $$renderer4.push(`<div class="text-rc-red-600">Invalid test: expected must be the last block (“Agent: …”), with a prompt
+													“User: …” block right above it</div>`);
+                } else {
+                  $$renderer4.push("<!--[!-->");
+                }
+                $$renderer4.push(`<!--]--> <input class="card-input-subtle svelte-rdft55" type="text" placeholder="Author"${attr("value", t.author)}/> <div class="flex items-center gap-2">`);
                 $$renderer4.select(
                   {
                     class: "rc-scroll rc-input w-36",
@@ -2133,65 +2154,86 @@ function _page($$renderer, $$props) {
                     }
                     $$renderer5.push(`<!--]-->`);
                   },
-                  "svelte-f5z0gs"
+                  "svelte-rdft55"
                 );
-                $$renderer4.push(` <input class="rc-input min-w-0 flex-1 svelte-f5z0gs" type="text"${attr("value", t.assertion_value)} placeholder="Assertion value"/> <button title="Duplicate" class="mr-2 shrink-0 text-gray-300 hover:text-gray-100 disabled:cursor-not-allowed disabled:opacity-50"${attr("disabled", running || isSaving, true)}>`);
-                Icon($$renderer4, { icon: "mdi:content-copy", width: 18, height: 18 });
-                $$renderer4.push(`<!----></button></div> <div class="flex items-center gap-3"><div class="flex items-center gap-1"><span class="text-xs uppercase tracking-wide text-gray-400">Timeout</span> <input class="rc-input no-spin w-16 svelte-f5z0gs" type="number" min="0" max="999"${attr("value", t.timeout)} placeholder="Timeout"/> <span class="text-xs text-gray-400">sec</span></div> <div class="flex items-center gap-1"><span class="text-xs uppercase tracking-wide text-gray-400">Similarity</span> <input class="rc-input no-spin w-16 svelte-f5z0gs" type="number" min="0" max="100"${attr("value", t.similarity_threshold)} placeholder="Similarity (%)"/> <span class="text-xs text-gray-400">%</span></div></div> <div class="text-xs uppercase tracking-wide text-gray-400">Test context</div> `);
-                ContextEditor($$renderer4, {
-                  class: "h-24",
-                  context: parseJson(t.context),
-                  suppressErrors: true,
-                  defaultFormat: "yaml",
-                  hideHeader: true,
-                  onContextChanged: (ctx) => {
-                    try {
-                      const str = JSON.stringify(ctx);
-                      t.context = str;
-                      saveRowDebounced(t.id, { context: ctx });
-                    } catch {
-                    }
-                  }
+                $$renderer4.push(` <input class="rc-input min-w-0 flex-1 svelte-rdft55" type="text"${attr("value", t.assertion_value)} placeholder="Assertion value"/></div> <div class="flex flex-wrap items-center gap-3"><div class="flex items-center gap-2"><span class="text-rc-gray-300">Timeout (s)</span> <input class="rc-input no-spin w-16 svelte-rdft55" type="number" min="0" max="999"${attr("value", t.timeout)} placeholder="Timeout"/></div> <div class="flex items-center gap-2"><span class="text-rc-gray-300">Similarity (%)</span> <input class="rc-input no-spin w-16 svelte-rdft55" type="number" min="0" max="100"${attr("value", t.similarity_threshold)} placeholder="Similarity (%)"/></div> `);
+                Button($$renderer4, {
+                  variant: "secondary",
+                  size: "small",
+                  disabled: running || isSaving,
+                  onclick: () => openTestContext(t),
+                  children: ($$renderer5) => {
+                    $$renderer5.push(`<!---->Edit context`);
+                  },
+                  $$slots: { default: true }
                 });
-                $$renderer4.push(`<!----> <div class="text-xs uppercase tracking-wide text-gray-400">Conversation</div> `);
+                $$renderer4.push(`<!----></div> `);
                 MiniEditor($$renderer4, {
-                  value: t.conversation,
-                  onChange: (v) => t.conversation = v,
-                  onBlur: () => saveRow(t.id, { conversation: t.conversation }),
-                  language: "markdown",
-                  height: 140
+                  value: t.conversation_ui,
+                  highlights: expectedInfo.highlight ? [
+                    {
+                      startLine: expectedInfo.highlight.startLine,
+                      endLine: expectedInfo.highlight.endLine,
+                      className: "test-expected-output"
+                    }
+                  ] : void 0,
+                  height: 220
                 });
-                $$renderer4.push(`<!----> <div class="text-xs uppercase tracking-wide text-gray-400">Expected</div> `);
-                MiniEditor($$renderer4, {
-                  value: t.expected_output,
-                  onChange: (v) => t.expected_output = v,
-                  onBlur: () => saveRow(t.id, { expected_output: t.expected_output }),
-                  language: "markdown",
-                  height: 120
+                $$renderer4.push(`<!----></div> <div class="rounded-b border-t border-rc-gray-700 bg-rc-gray-980 p-2"><div class="flex flex-wrap items-center gap-3">`);
+                Button($$renderer4, {
+                  size: "small",
+                  disabled: running || isSaving || !expectedInfo.valid,
+                  title: !expectedInfo.valid ? "Expected must be the last block (“Agent: …”) and there must be a prompt “User: …” block right above it" : "Run",
+                  onclick: () => runSingle(t.id),
+                  children: ($$renderer5) => {
+                    $$renderer5.push(`<!---->▶ Run`);
+                  },
+                  $$slots: { default: true }
                 });
-                $$renderer4.push(`<!----></div></div> <div class="md:col-span-8"><div class="mb-2 ml-3 text-sm">`);
+                $$renderer4.push(`<!----> `);
                 if (t.result) {
                   $$renderer4.push("<!--[-->");
-                  $$renderer4.push(`<div class="text-lg font-extrabold"><span class="text-rc-gray-200">${escape_html(t.result.runs.filter((r) => r.success).length)}/${escape_html(t.result.total_runs)}</span> `);
+                  const trend = getExpectedTrend(t.result, runExpectedRuns);
+                  $$renderer4.push(`<div class="text-lg font-extrabold"><span class="text-rc-gray-200">${escape_html(t.result.successes)}/${escape_html(t.result.total_runs)}</span> `);
+                  if (trend?.showExpected) {
+                    $$renderer4.push("<!--[-->");
+                    $$renderer4.push(`<span class="ml-2 text-rc-gray-400">(${escape_html(trend.expectedSuccesses)}/${escape_html(trend.expectedRuns)} expected)</span>`);
+                  } else {
+                    $$renderer4.push("<!--[!-->");
+                  }
+                  $$renderer4.push(`<!--]--> `);
+                  if (trend?.significant) {
+                    $$renderer4.push("<!--[-->");
+                    $$renderer4.push(`<span${attr_class(`ml-2 whitespace-nowrap font-semibold ${trend.direction === "higher" ? "text-rc-green-200" : "text-rc-red-600"}`, "svelte-rdft55")}>`);
+                    Icon$1($$renderer4, {
+                      name: trend.direction === "higher" ? "chevron-up" : "chevron-down",
+                      width: 18,
+                      height: 18,
+                      class: "inline-block align-text-bottom"
+                    });
+                    $$renderer4.push(`<!----><span class="ml-1">significant</span></span>`);
+                  } else {
+                    $$renderer4.push("<!--[!-->");
+                  }
+                  $$renderer4.push(`<!--]--> `);
                   if (t.result.isOld) {
                     $$renderer4.push("<!--[-->");
                     $$renderer4.push(`<span class="text-rc-gray-400">(old)</span>`);
                   } else {
                     $$renderer4.push("<!--[!-->");
                   }
-                  $$renderer4.push(`<!--]--></div> <span class="block font-medium text-rc-gray-400">Success:</span>`);
+                  $$renderer4.push(`<!--]--></div>`);
                 } else {
                   $$renderer4.push("<!--[!-->");
-                  $$renderer4.push(`<span class="ml-3 text-gray-500">—</span>`);
                 }
                 $$renderer4.push(`<!--]--></div> `);
                 if (t.result) {
                   $$renderer4.push("<!--[-->");
-                  $$renderer4.push(`<div class="flex flex-col gap-3"><!--[-->`);
+                  $$renderer4.push(`<div class="mt-3 flex flex-col gap-4"><!--[-->`);
                   const each_array_2 = ensure_array_like(t.result.runs);
                   for (let $$index_1 = 0, $$length2 = each_array_2.length; $$index_1 < $$length2; $$index_1++) {
                     let run2 = each_array_2[$$index_1];
-                    $$renderer4.push(`<div class="max-w-3xl rounded-md border border-rc-gray-700 bg-rc-gray-800/40 px-3 py-2"><div${attr_class(`w-full whitespace-pre-wrap break-words rounded-lg px-3 py-2 text-rc-gray-100 ${run2.success ? "bg-rc-green-700 backdrop-blur-md" : ""}`, "svelte-f5z0gs")}${attr_style(!run2.success ? "background-color: rgb(156 80 80)" : "")}>${escape_html(run2.text || "—")}</div> <div${attr_class(`mt-2 flex w-full flex-col gap-x-4 whitespace-pre-wrap break-words text-sm ${run2.success ? "text-rc-green-200" : "text-rc-red-800"}`, "svelte-f5z0gs")}><div class="flex items-center gap-0">`);
+                    $$renderer4.push(`<div class="flex flex-col gap-2"><div${attr_class(`w-full whitespace-pre-wrap break-words rounded-lg px-3 py-2 text-rc-gray-100 ${run2.success ? "bg-rc-green-700 backdrop-blur-md" : ""}`, "svelte-rdft55")}${attr_style(!run2.success ? "background-color: rgb(156 80 80)" : "")}>${escape_html(run2.text || "—")}</div> <div${attr_class(`flex w-full flex-col gap-x-4 whitespace-pre-wrap break-words ${run2.success ? "text-rc-green-200" : "text-rc-red-800"}`, "svelte-rdft55")}><div class="flex items-center gap-0">`);
                     if (run2.success) {
                       $$renderer4.push("<!--[-->");
                       Icon$1($$renderer4, {
@@ -2209,90 +2251,50 @@ function _page($$renderer, $$props) {
                         class: "shrink-0 scale-110 self-start"
                       });
                     }
-                    $$renderer4.push(`<!--]--> <span>${escape_html(run2.success ? "Pass" : formatEvaluation(run2.evaluation) || "Fail")}</span></div> <div class="mt-1 flex flex-wrap items-center gap-x-4">`);
-                    Button($$renderer4, {
-                      variant: "outlined",
-                      size: "small",
-                      class: "border-none !px-0 text-sm text-rc-gray-400 underline decoration-dashed underline-offset-4 hover:bg-transparent hover:text-rc-gray-200",
-                      onclick: () => openViewer("Rendered Prompt", run2.links.renderedPrompt),
-                      children: ($$renderer5) => {
-                        $$renderer5.push(`<!---->Rendered prompt`);
-                      },
-                      $$slots: { default: true }
-                    });
-                    $$renderer4.push(`<!----> `);
-                    Button($$renderer4, {
-                      variant: "outlined",
-                      size: "small",
-                      class: "border-none !px-0 text-sm text-rc-gray-400 underline decoration-dashed underline-offset-4 hover:bg-transparent hover:text-rc-gray-200",
-                      onclick: () => openViewer("Received Prompt", run2.links.receivedPrompt),
-                      children: ($$renderer5) => {
-                        $$renderer5.push(`<!---->Received prompt`);
-                      },
-                      $$slots: { default: true }
-                    });
-                    $$renderer4.push(`<!----> `);
-                    Button($$renderer4, {
-                      variant: "outlined",
-                      size: "small",
-                      class: "border-none !px-0 text-sm text-rc-gray-400 underline decoration-dashed underline-offset-4 hover:bg-transparent hover:text-rc-gray-200",
-                      onclick: () => openViewer("Conversation", run2.links.conversation),
-                      children: ($$renderer5) => {
-                        $$renderer5.push(`<!---->Conversation`);
-                      },
-                      $$slots: { default: true }
-                    });
-                    $$renderer4.push(`<!----> `);
-                    Button($$renderer4, {
-                      variant: "outlined",
-                      size: "small",
-                      class: "border-none !px-0 text-sm text-rc-gray-400 underline decoration-dashed underline-offset-4 hover:bg-transparent hover:text-rc-gray-200",
-                      onclick: () => openViewer("Context", run2.links.context),
-                      children: ($$renderer5) => {
-                        $$renderer5.push(`<!---->Context`);
-                      },
-                      $$slots: { default: true }
-                    });
-                    $$renderer4.push(`<!----> `);
-                    Button($$renderer4, {
-                      variant: "outlined",
-                      size: "small",
-                      class: "border-none !px-0 text-sm text-rc-gray-400 underline decoration-dashed underline-offset-4 hover:bg-transparent hover:text-rc-gray-200",
-                      onclick: () => openViewer("Logs", run2.links.logs),
-                      children: ($$renderer5) => {
-                        $$renderer5.push(`<!---->Logs`);
-                      },
-                      $$slots: { default: true }
-                    });
-                    $$renderer4.push(`<!----> `);
-                    Button($$renderer4, {
-                      variant: "outlined",
-                      size: "small",
-                      class: "border-none !px-0 text-sm text-rc-gray-400 underline decoration-dashed underline-offset-4 hover:bg-transparent hover:text-rc-gray-200",
-                      onclick: () => openViewer("Traces", run2.links.traces),
-                      children: ($$renderer5) => {
-                        $$renderer5.push(`<!---->Traces`);
-                      },
-                      $$slots: { default: true }
-                    });
-                    $$renderer4.push(`<!----></div></div></div>`);
+                    $$renderer4.push(`<!--]--> <span>${escape_html(run2.success ? "Pass" : formatEvaluation(run2.evaluation) || "Fail")}</span></div> <div class="mt-1 flex flex-wrap items-center gap-x-4"><button type="button" class="pseudolink svelte-rdft55">Rendered prompt</button> <button type="button" class="pseudolink svelte-rdft55">Received prompt</button> <button type="button" class="pseudolink svelte-rdft55">Conversation</button> <button type="button" class="pseudolink svelte-rdft55">Context</button> <button type="button" class="pseudolink svelte-rdft55">Logs</button> <button type="button" class="pseudolink svelte-rdft55">Traces</button> <button type="button" class="pseudolink svelte-rdft55"${attr("disabled", analyzeRunning, true)}>Analyze</button></div></div></div>`);
                   }
                   $$renderer4.push(`<!--]--></div>`);
                 } else {
                   $$renderer4.push("<!--[!-->");
-                  $$renderer4.push(`<span class="ml-6 text-gray-500">—</span>`);
                 }
-                $$renderer4.push(`<!--]--></div></div></div></td></tr>`);
+                $$renderer4.push(`<!--]--></div></div></div></div></div>`);
               }
-              $$renderer4.push(`<!--]--></tbody></table></div>`);
+              $$renderer4.push(`<!--]--></div>`);
             }
-          });
-        } else {
-          $$renderer3.push("<!--[!-->");
+            $$renderer4.push(`<!--]-->`);
+          }
+          $$renderer4.push(`<!--]--></div> `);
+          if (testContextPopoverTestId) {
+            $$renderer4.push("<!--[-->");
+            const row = tests.find((t) => t.id === testContextPopoverTestId);
+            {
+              let children = function($$renderer5) {
+                $$renderer5.push(`<div class="min-h-0 flex-1 overflow-hidden p-3"><!---->`);
+                {
+                  ContextEditor($$renderer5, {
+                    class: "h-full min-h-0 border-thin border-rc-gray-650 shadow-inner-white-soft backdrop-blur-3xl",
+                    context: testContextDraft,
+                    suppressErrors: true,
+                    defaultFormat: "yaml",
+                    hideHeader: true,
+                    onContextChanged: (ctx) => testContextDraft = ctx
+                  });
+                }
+                $$renderer5.push(`<!----></div>`);
+              };
+              PopoverModal($$renderer4, {
+                title: `Test context — ${row?.scenario || "Test"}`,
+                onClose: () => void closeTestContextPopover(),
+                children
+              });
+            }
+          } else {
+            $$renderer4.push("<!--[!-->");
+          }
+          $$renderer4.push(`<!--]-->`);
         }
-        $$renderer3.push(`<!--]-->`);
-      }
-      $$renderer3.push(`<!--]--></div></div></div>`);
+      });
+      $$renderer3.push(`<!----></div></div>`);
     }
     do {
       $$settled = true;
@@ -2304,4 +2306,4 @@ function _page($$renderer, $$props) {
 }
 
 export { _page as default };
-//# sourceMappingURL=_page.svelte-N6xQwQzn.js.map
+//# sourceMappingURL=_page.svelte-C1YaC_Qg.js.map