@recombine-ai/bosun 0.16.2 → 0.18.0

package/build/server/chunks/{_page.svelte-BSVJ2d2G.js → _page.svelte-nQ_thh1S.js}

@@ -1,17 +1,17 @@
-import { w as head, p as ensure_array_like, o as attr_class, k as escape_html, r as clsx, m as attr, x as attr_style, l as derived, q as attributes } from './index-LMsd1ih2.js';
-import { B as Button } from './Button-DH7gdU0s.js';
-import { M as Modal } from './Modal-CGJuNpZ0.js';
-import { I as Input } from './Input-CUJ9Cg7b.js';
-import { o as onDestroy } from './index-server-BRr5mZEq.js';
-import { I as Icon$1, h as html } from './Icon-DP1YKCRV.js';
-import './client-DO-hqMbT.js';
+import { x as head, p as ensure_array_like, o as attr_class, k as escape_html, t as clsx, m as attr, y as attr_style, l as derived, q as attributes } from './index-Do08hKbr.js';
+import { B as Button } from './Button-Huqrc2wY.js';
+import { M as Modal } from './Modal-B_fVcSLK.js';
+import { I as Input } from './Input-D-uxKbuX.js';
+import { o as onDestroy } from './index-server-CHFC9Apw.js';
+import { I as Icon$1, h as html } from './Icon-BJfQMgGu.js';
+import './client-D1Y2ZQJN.js';
 import 'js-yaml';
 import './trpc-client-D8L6vWf4.js';
-import { g as getEditWorkspaceContext } from './edit-workspace-context-tyHBPujC.js';
+import { g as getEditWorkspaceContext } from './edit-workspace-context-ClnqepZ4.js';
 import { i as isMainStepName } from './steps-DzZCg39h.js';
-import './root-4q2Pr7ck.js';
-import './state.svelte-Bd1i-u49.js';
-import './index3-BdRUAffj.js';
+import './root-DoSdrdfw.js';
+import './state.svelte-B2OaeWMp.js';
+import './index3-C9gC86lU.js';
 import './shared-server-DaWdgxVh.js';
 import '@trpc/client';
 
@@ -1389,6 +1389,7 @@ function MiniEditor($$renderer, $$props) {
     let {
       height = 120
     } = $$props;
+    let editorDecorations = null;
     function handleWheelCapture(e) {
       e.stopPropagation();
     }
@@ -1404,6 +1405,8 @@ function MiniEditor($$renderer, $$props) {
         }
       }
       disposables = [];
+      editorDecorations?.dispose();
+      editorDecorations = null;
       const toDisposeModel = model;
       model = null;
       try {
@@ -1424,7 +1427,7 @@ function MiniEditor($$renderer, $$props) {
     });
     $$renderer2.push(`<div class="w-full"${attr_style(
       // Update language when prop changes without recreating the editor
-      // External value changes should update the model (avoid loops)
+      // Keep the model and its semantic decorations in sync with external changes.
       `height:${height}px`
     )}></div>`);
   });
@@ -1436,7 +1439,9 @@ function RunResultActions($$renderer, $$props) {
       actionsInert = false,
       addTestDisabled = false,
       showAnalyze = true,
-      analyzeDisabled = false
+      analyzeDisabled = false,
+      fixDisabled = false,
+      onFix
     } = $$props;
     let addStatus = "idle";
     onDestroy(() => {
@@ -1444,7 +1449,14 @@ function RunResultActions($$renderer, $$props) {
     let $$settled = true;
     let $$inner_renderer;
     function $$render_inner($$renderer3) {
-      $$renderer3.push(`<div class="mt-1 flex flex-wrap items-center gap-x-2 gap-y-1"><button type="button"${attr("inert", actionsInert, true)}${attr("disabled", actionsDisabled || addTestDisabled || addStatus === "adding", true)} class="inline-flex items-center text-nowrap rounded border-0 bg-transparent px-1 py-0.5 text-rc-gray-400 transition-colors hover:bg-rc-gray-900 hover:text-rc-gray-200 focus:outline-none focus-visible:outline-none disabled:cursor-not-allowed disabled:opacity-50 disabled:hover:bg-transparent disabled:hover:text-rc-gray-400">${escape_html("Add test")}</button> `);
+      $$renderer3.push(`<div class="mt-1 flex flex-wrap items-center gap-x-2 gap-y-1">`);
+      if (onFix) {
+        $$renderer3.push("<!--[0-->");
+        $$renderer3.push(`<button type="button"${attr("inert", actionsInert, true)}${attr("disabled", actionsDisabled || fixDisabled, true)} class="inline-flex items-center text-nowrap rounded border-0 bg-transparent px-1 py-0.5 text-rc-gray-400 transition-colors hover:bg-rc-gray-900 hover:text-rc-gray-200 focus:outline-none focus-visible:outline-none disabled:cursor-not-allowed disabled:opacity-50 disabled:hover:bg-transparent disabled:hover:text-rc-gray-400">Fix</button>`);
+      } else {
+        $$renderer3.push("<!--[-1-->");
+      }
+      $$renderer3.push(`<!--]--> <button type="button"${attr("inert", actionsInert, true)}${attr("disabled", actionsDisabled || addTestDisabled || addStatus === "adding", true)} class="inline-flex items-center text-nowrap rounded border-0 bg-transparent px-1 py-0.5 text-rc-gray-400 transition-colors hover:bg-rc-gray-900 hover:text-rc-gray-200 focus:outline-none focus-visible:outline-none disabled:cursor-not-allowed disabled:opacity-50 disabled:hover:bg-transparent disabled:hover:text-rc-gray-400">${escape_html("Add test")}</button> `);
       if (showAnalyze) {
         $$renderer3.push("<!--[0-->");
         $$renderer3.push(`<button type="button"${attr("disabled", analyzeDisabled, true)} class="inline-flex items-center text-nowrap rounded border-0 bg-transparent px-1 py-0.5 text-rc-gray-400 transition-colors hover:bg-rc-gray-900 hover:text-rc-gray-200 focus:outline-none focus-visible:outline-none disabled:cursor-not-allowed disabled:opacity-50 disabled:hover:bg-transparent disabled:hover:text-rc-gray-400">Analyze</button>`);
@@ -1473,6 +1485,29 @@ function RunResultActions($$renderer, $$props) {
     $$renderer2.subsume($$inner_renderer);
   });
 }
+function formatTestFixAssertion(label, value) {
+  const normalizedLabel = label.trim() === "LLM" ? "" : label.trim();
+  return [normalizedLabel, value.trim()].filter(Boolean).join(" ");
+}
+function buildTestFixRequest(input) {
+  const assertion = formatTestFixAssertion(input.assertionLabel, input.assertionValue);
+  return `You will be given the context, the conversation and the desired output and assertion about it. First, find why the prompt is not producing the correct result. It's usually one concrete instruction within the relevant section, or a contradiction. Then, change the prompt minimally so this will work, but keeping the prompt clean and comprehensive is more important. Try to provide as general solution as possible. Try to change just one part of the prompt, not multiple parts. Avoid adding exceptions and overrides. Avoid duplicating information that already exists in the prompt.
+
+Context:
+${input.context.trim()}
+
+Conversation:
+${input.conversation.trim()}
+
+Desired outcome:
+${input.expectedOutput.trim()}
+
+Assertion about outcome:
+${assertion}
+
+Current incorrect outcome:
+${input.currentIncorrectOutcome.trim()}`;
+}
 function erf(x) {
   const sign = x < 0 ? -1 : 1;
   const ax = Math.abs(x);
@@ -1538,6 +1573,70 @@ function getExpectedTrend(result, expectedRuns) {
     comparisonLabel
   };
 }
+const baseAssertionTypes = [
+  { value: "llm-rubric", label: "LLM" },
+  { value: "answer-relevance", label: "Answer Relevance (LLM)" },
+  { value: "context-faithfulness", label: "Context Faithfulness (LLM)" },
+  { value: "context-recall", label: "Context Recall (LLM)" },
+  { value: "context-relevance", label: "Context Relevance (LLM)" },
+  { value: "model-graded-closedqa", label: "Model-Graded Closed QA" },
+  { value: "model-graded-factuality", label: "Model-Graded Factuality" },
+  { value: "moderation", label: "Moderation" },
+  { value: "classifier", label: "Classifier" },
+  { value: "g-eval", label: "G-Eval" },
+  { value: "factuality", label: "Factuality" },
+  { value: "conversation-relevance", label: "Conversation Relevance" },
+  { value: "equals", label: "Equals" },
+  { value: "contains", label: "Contains" },
+  { value: "icontains", label: "Contains (case-insensitive)" },
+  { value: "contains-all", label: "Contains All" },
+  { value: "contains-any", label: "Contains Any" },
+  { value: "icontains-all", label: "Contains All (iCase)" },
+  { value: "icontains-any", label: "Contains Any (iCase)" },
+  { value: "starts-with", label: "Starts With" },
+  { value: "regex", label: "Regex" },
+  { value: "levenshtein", label: "Levenshtein Distance" },
+  { value: "is-refusal", label: "Is Refusal" },
+  { value: "is-html", label: "Is HTML" },
+  { value: "contains-html", label: "Contains HTML" },
+  { value: "is-json", label: "Is JSON" },
+  { value: "contains-json", label: "Contains JSON" },
+  { value: "is-xml", label: "Is XML" },
+  { value: "contains-xml", label: "Contains XML" },
+  { value: "is-sql", label: "Is SQL" },
+  { value: "contains-sql", label: "Contains SQL" },
+  { value: "similar", label: "Semantic Similarity" },
+  { value: "rouge-n", label: "ROUGE-N" },
+  { value: "bleu", label: "BLEU" },
+  { value: "gleu", label: "GLEU" },
+  { value: "meteor", label: "METEOR" },
+  { value: "perplexity", label: "Perplexity" },
+  { value: "perplexity-score", label: "Perplexity Score" },
+  { 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" },
+  { value: "is-valid-function-call", label: "Is Valid Function Call" },
+  { value: "is-valid-openai-function-call", label: "Is Valid OpenAI Function Call" },
+  { value: "is-valid-openai-tools-call", label: "Is Valid OpenAI Tools Call" },
+  { value: "guardrails", label: "Guardrails" },
+  { value: "javascript", label: "JavaScript (custom)" },
+  { value: "python", label: "Python (custom)" },
+  { value: "webhook", label: "Webhook (custom)" },
+  { value: "pi", label: "Pi Scorer" }
+];
+const assertionTypes = [
+  ...baseAssertionTypes,
+  ...baseAssertionTypes.map((option) => ({
+    value: `not-${option.value}`,
+    label: `Not ${option.label}`
+  }))
+];
+function assertionLabel(value) {
+  return assertionTypes.find((option) => option.value === value)?.label ?? value;
+}
 function _page($$renderer, $$props) {
   $$renderer.component(($$renderer2) => {
     const { state: workspace, actions } = getEditWorkspaceContext();
@@ -1569,89 +1668,18 @@ function _page($$renderer, $$props) {
       showViewerModal = true;
     }
     let analyzeRunning = false;
-    const baseAssertionTypes = [
-      { value: "llm-rubric", label: "LLM" },
-      { value: "answer-relevance", label: "Answer Relevance (LLM)" },
-      {
-        value: "context-faithfulness",
-        label: "Context Faithfulness (LLM)"
-      },
-      { value: "context-recall", label: "Context Recall (LLM)" },
-      { value: "context-relevance", label: "Context Relevance (LLM)" },
-      {
-        value: "model-graded-closedqa",
-        label: "Model-Graded Closed QA"
-      },
-      {
-        value: "model-graded-factuality",
-        label: "Model-Graded Factuality"
-      },
-      { value: "moderation", label: "Moderation" },
-      { value: "classifier", label: "Classifier" },
-      { value: "g-eval", label: "G-Eval" },
-      { value: "factuality", label: "Factuality" },
-      {
-        value: "conversation-relevance",
-        label: "Conversation Relevance"
-      },
-      { value: "equals", label: "Equals" },
-      { value: "contains", label: "Contains" },
-      { value: "icontains", label: "Contains (case-insensitive)" },
-      { value: "contains-all", label: "Contains All" },
-      { value: "contains-any", label: "Contains Any" },
-      { value: "icontains-all", label: "Contains All (iCase)" },
-      { value: "icontains-any", label: "Contains Any (iCase)" },
-      { value: "starts-with", label: "Starts With" },
-      { value: "regex", label: "Regex" },
-      { value: "levenshtein", label: "Levenshtein Distance" },
-      { value: "is-refusal", label: "Is Refusal" },
-      { value: "is-html", label: "Is HTML" },
-      { value: "contains-html", label: "Contains HTML" },
-      { value: "is-json", label: "Is JSON" },
-      { value: "contains-json", label: "Contains JSON" },
-      { value: "is-xml", label: "Is XML" },
-      { value: "contains-xml", label: "Contains XML" },
-      { value: "is-sql", label: "Is SQL" },
-      { value: "contains-sql", label: "Contains SQL" },
-      { value: "similar", label: "Semantic Similarity" },
-      { value: "rouge-n", label: "ROUGE-N" },
-      { value: "bleu", label: "BLEU" },
-      { value: "gleu", label: "GLEU" },
-      { value: "meteor", label: "METEOR" },
-      { value: "perplexity", label: "Perplexity" },
-      { value: "perplexity-score", label: "Perplexity Score" },
-      { 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" },
-      {
-        value: "is-valid-function-call",
-        label: "Is Valid Function Call"
-      },
-      {
-        value: "is-valid-openai-function-call",
-        label: "Is Valid OpenAI Function Call"
-      },
-      {
-        value: "is-valid-openai-tools-call",
-        label: "Is Valid OpenAI Tools Call"
-      },
-      { value: "guardrails", label: "Guardrails" },
-      { value: "javascript", label: "JavaScript (custom)" },
-      { value: "python", label: "Python (custom)" },
-      { value: "webhook", label: "Webhook (custom)" },
-      { value: "pi", label: "Pi Scorer" }
-    ];
-    const assertionTypes = [
-      ...baseAssertionTypes,
-      ...baseAssertionTypes.map((opt) => ({ value: `not-${opt.value}`, label: `Not ${opt.label}` }))
-    ];
+    let tests = [];
     let running = false;
     Promise.resolve();
     let pendingSaveCount = 0;
     const isSaving = derived(() => pendingSaveCount > 0);
+    let expandedTestIds = [];
+    function historyEditKey(testId) {
+      return `${agent()}\0${region() ?? ""}\0${testId}`;
+    }
+    function pendingHistoryEdit(testId) {
+      return workspace.testHistoryEdits[historyEditKey(testId)] ?? null;
+    }
     function parseHistoryTest(timestamp, row) {
       const test = row.test;
       const id = test.id;
@@ -1718,8 +1746,43 @@ function _page($$renderer, $$props) {
       }
       return { casesPassing, casesTotal, attemptsPassing, attemptsTotal };
     }
-    function assertionLabel(value) {
-      return assertionTypes.find((opt) => opt.value === value)?.label ?? value;
+    function testAiEditTarget(row, kind, currentIncorrectOutcome) {
+      const selectedMainStep = isMainStepName(row.step) && isMainStepName(workspace.currentStep?.name) ? workspace.currentStep : null;
+      const exactStep = workspace.steps.find((step2) => step2.name === row.step);
+      const promptStep = selectedMainStep ?? exactStep ?? workspace.steps.find((step2) => isMainStepName(row.step) ? isMainStepName(step2.name) : false);
+      const step = promptStep?.name ?? row.step;
+      const persistedPrompt = promptStep?.prompt ?? "";
+      const promptDraft = workspace.promptDrafts[step] ?? persistedPrompt;
+      const promptProposal = workspace.promptEdits[`${agent()}\0${step}`];
+      return {
+        kind,
+        testId: row.id,
+        step,
+        prompt: promptProposal?.suggestedPrompt ?? promptDraft,
+        basePrompt: kind === "test-fix" ? persistedPrompt : promptProposal?.originalPrompt ?? promptDraft,
+        conversation: row.conversation,
+        expectedOutput: row.expected_output,
+        context: row.context,
+        assertionLabel: row.evaluation_mode === "json-deep-equal" ? "" : assertionLabel(row.assertion_type),
+        assertionValue: row.evaluation_mode === "json-deep-equal" ? "" : row.assertion_value,
+        currentIncorrectOutcome
+      };
+    }
+    function openTestFix(row, currentIncorrectOutcome) {
+      showHistoryModal = false;
+      const currentRow = tests.find((test) => test.id === row.id) ?? row;
+      void setTestExpanded(currentRow.id, true);
+      actions.openAiEdit(
+        buildTestFixRequest({
+          context: currentRow.context,
+          conversation: currentRow.conversation,
+          expectedOutput: currentRow.expected_output,
+          assertionLabel: currentRow.evaluation_mode === "json-deep-equal" ? "" : assertionLabel(currentRow.assertion_type),
+          assertionValue: currentRow.evaluation_mode === "json-deep-equal" ? "" : currentRow.assertion_value,
+          currentIncorrectOutcome
+        }),
+        testAiEditTarget(currentRow, "test-fix", currentIncorrectOutcome)
+      );
     }
     function formatRunDate(timestamp) {
       const d = new Date(timestamp);
@@ -1839,9 +1902,32 @@ ${editorExpected}
         }
       };
     }
+    function buildExpectedOutputHighlights(row, highlight) {
+      if (!highlight) return [];
+      return [
+        {
+          startLine: highlight.startLine,
+          endLine: highlight.endLine,
+          className: row.evaluation_mode === "json-deep-equal" ? "test-expected-json-output" : "test-expected-output"
+        }
+      ];
+    }
     function resultCountClass(result) {
       return result.total_runs > 0 && result.successes < result.total_runs ? "text-rc-red-600" : "text-rc-gray-200";
     }
+    function setTestExpanded(id, expanded) {
+      const agentId = (agent() ?? "").trim();
+      if (!agentId) return Promise.resolve();
+      const normalizedId = String(id ?? "").trim();
+      if (!normalizedId) return Promise.resolve();
+      const next = new Set(expandedTestIds);
+      const has = next.has(normalizedId);
+      if (expanded === has) return Promise.resolve();
+      next.add(normalizedId);
+      const nextIds = Array.from(next);
+      expandedTestIds = nextIds;
+      return Promise.resolve();
+    }
     onDestroy(() => {
     });
     let $$settled = true;
@@ -1980,13 +2066,7 @@ ${editorExpected}
                   $$renderer4.push(`<!----></div> `);
                   MiniEditor($$renderer4, {
                     value: t.conversation_ui,
-                    highlights: expectedInfo.highlight ? [
-                      {
-                        startLine: expectedInfo.highlight.startLine,
-                        endLine: expectedInfo.highlight.endLine,
-                        className: t.evaluation_mode === "json-deep-equal" ? "test-expected-json-output" : "test-expected-output"
-                      }
-                    ] : void 0,
+                    highlights: buildExpectedOutputHighlights(t, expectedInfo.highlight),
                     height: 220
                   });
                   $$renderer4.push(`<!----></div> <div class="rounded-b border-t border-rc-gray-700 bg-rc-gray-980 p-2">`);
@@ -2033,7 +2113,9 @@ ${editorExpected}
                         actionsInert: isSaving(),
                         addTestDisabled: !run.text.trim(),
                         showAnalyze: t.evaluation_mode !== "json-deep-equal",
-                        analyzeDisabled: analyzeRunning
+                        analyzeDisabled: analyzeRunning,
+                        fixDisabled: !!actions.getActivePromptEditId(),
+                        onFix: !run.success && !pendingHistoryEdit(t.id) ? () => openTestFix(t, run.text || run.errors || "—") : void 0
                       });
                       $$renderer4.push(`<!----></div></div>`);
                     }
@@ -2126,4 +2208,4 @@ ${editorExpected}
 }
 
 export { _page as default };
-//# sourceMappingURL=_page.svelte-BSVJ2d2G.js.map
+//# sourceMappingURL=_page.svelte-nQ_thh1S.js.map