@recombine-ai/platform 0.8.0-test.2 → 0.8.1

package/build/bosun-app/backend/server.js

@@ -34571,16 +34571,18 @@ function createAgents(dispatcher, statePersistence, contextPersistence) {
   let basePath = "";
   let currentAgent = "";
   let currentRegion = null;
+  let currentVariant = null;
   let currentStateId = null;
   let lastLoadedImportNonce = null;
-  function getAgentCacheKey(agentId, region) {
-    return region ? `${agentId}.${region}` : agentId;
+  function getAgentCacheKey(agentId, region, variant = null) {
+    const key = region ? `${agentId}.${region}` : agentId;
+    return variant ? `${key}.${variant}` : key;
   }
   async function getCurrentAgent() {
     if (!currentAgent) {
       currentAgent = Object.keys(factories)[0];
     }
-    const cacheKey = getAgentCacheKey(currentAgent, currentRegion);
+    const cacheKey = getAgentCacheKey(currentAgent, currentRegion, currentVariant);
     let agent = agents[cacheKey];
     if (!agent) {
       const factory = factories[currentAgent];
@@ -34594,6 +34596,7 @@ function createAgents(dispatcher, statePersistence, contextPersistence) {
         ...dispatcher,
         agentId: cacheKey,
         region: currentRegion,
+        variantId: currentVariant,
         contextPersistence,
         statePersistence,
         factory,
@@ -34623,12 +34626,12 @@ function createAgents(dispatcher, statePersistence, contextPersistence) {
       await api.load();
       const agentId = currentAgent || Object.keys(factories)[0];
       if (!agentId) return;
-      const cacheKey = getAgentCacheKey(agentId, currentRegion);
+      const cacheKey = getAgentCacheKey(agentId, currentRegion, currentVariant);
       delete agents[cacheKey];
       delete agentStates[cacheKey];
-      await api.select(agentId, currentStateId, currentRegion);
+      await api.select(agentId, currentStateId, currentRegion, {}, currentVariant);
     },
-    select: async (agentId, stateId, region, promptOverrides = {}) => {
+    select: async (agentId, stateId, region, promptOverrides = {}, variant) => {
       console.log(
         "createAgents, select; agents: ",
         Object.keys(agents),
@@ -34637,6 +34640,7 @@ function createAgents(dispatcher, statePersistence, contextPersistence) {
       );
       currentAgent = agentId;
       currentRegion = region || null;
+      currentVariant = variant || null;
       const agent = await getCurrentAgent();
       agent.setPromptOverrides({});
       await dispatcher.sendAgentSelected({
@@ -34695,8 +34699,8 @@ function createAgents(dispatcher, statePersistence, contextPersistence) {
     handleMessage: async (message) => {
       currentStateId = await (await getCurrentAgent()).handleMessage(message);
     },
-    hangup: async (agentId, region) => {
-      const cacheKey = getAgentCacheKey(agentId, region);
+    hangup: async (agentId, region, variant = null) => {
+      const cacheKey = getAgentCacheKey(agentId, region, variant);
       const agent = agents[cacheKey];
       if (!agent) {
         throw new Error(`Cannot hangup: cannot find agent: ${cacheKey}`);
@@ -34807,6 +34811,7 @@ async function loadAgent(props) {
     stepTracer,
     stepRegistry,
     logger,
+    variantId: props.variantId ?? void 0,
     apiCallTracer,
     conversationalTracer,
     scheduler,
@@ -35316,7 +35321,8 @@ async function oldServer(ws) {
                 msg.event.agent,
                 msg.event.state,
                 msg.event.region,
-                msg.event.promptOverrides
+                msg.event.promptOverrides,
+                msg.event.variant
               );
               break;
             case "prompt-overrides-updated":
@@ -35338,7 +35344,11 @@ async function oldServer(ws) {
               await agents.selectStep(msg.event.step);
               break;
             case "hangup":
-              await agents.hangup(msg.event.agentId, msg.event.region);
+              await agents.hangup(
+                msg.event.agentId,
+                msg.event.region,
+                msg.event.variant ?? null
+              );
               break;
             default:
               console.error(`Unknown event type: ${msg.event.type}`);
@@ -46164,7 +46174,7 @@ import { createContext as createContext3 } from "@recombine-ai/platform/bosun";
 import "@recombine-ai/platform/engine";
 import "@recombine-ai/platform/telescope";
 async function generateReplyVariants(props) {
-  const { agentId, region, states, count, context, promptOverrides = {} } = props;
+  const { agentId, region, variantId, states, count, context, promptOverrides = {} } = props;
   if (!Number.isFinite(count) || count <= 0) return [];
   const targetState = states.at(-1);
   if (!targetState) throw new Error("No saved roleplay state found for variants");
@@ -46187,6 +46197,7 @@ async function generateReplyVariants(props) {
         index,
         factory,
         region,
+        variantId,
         context,
         prevState,
         statesById,
@@ -46257,7 +46268,17 @@ function createVariantState(prev, index) {
   };
 }
 async function generateSingleReplyVariant(props) {
-  const { index, factory, region, context, prevState, statesById, message, promptOverrides } = props;
+  const {
+    index,
+    factory,
+    region,
+    variantId,
+    context,
+    prevState,
+    statesById,
+    message,
+    promptOverrides
+  } = props;
   const state = createVariantState(prevState?.id ?? null, index);
   const scheduleSeed = clone3(prevState?.schedule ?? {});
   const log = (entry) => {
@@ -46316,6 +46337,7 @@ async function generateSingleReplyVariant(props) {
     stepTracer,
     stepRegistry,
     logger,
+    variantId,
     apiCallTracer,
     conversationalTracer,
     scheduler,
@@ -46432,10 +46454,16 @@ function createRolePlayRouter(t3) {
       external_exports.object({
         agentId: AgentIdSchema,
         step: external_exports.string().trim().min(1).max(512),
-        prompt: external_exports.string().max(1e6)
+        prompt: external_exports.string().max(1e6),
+        variant: external_exports.string().trim().min(1).max(128).nullable().optional()
       })
     ).mutation(async ({ input }) => {
-      await persistPromptUpdate(input.agentId, input.step, input.prompt);
+      await persistPromptUpdate(
+        input.agentId,
+        input.step,
+        input.prompt,
+        input.variant ?? void 0
+      );
       return { ok: true };
     }),
     suggestPromptEdit: t3.procedure.input(
@@ -46443,10 +46471,16 @@ function createRolePlayRouter(t3) {
         agentId: AgentIdSchema,
         step: external_exports.string().trim().min(1).max(512),
         request: external_exports.string().trim().min(1).max(1e5),
-        progressId: external_exports.string().uuid()
+        progressId: external_exports.string().uuid(),
+        variant: external_exports.string().trim().min(1).max(128).nullable().optional()
       })
     ).mutation(async ({ input }) => {
-      const originalPrompt = await readEditablePrompt(input.agentId, input.step);
+      const variantId = input.variant ?? void 0;
+      const originalPrompt = await readEditablePrompt(
+        input.agentId,
+        input.step,
+        variantId
+      );
       const suggestedPrompt = await suggestPromptEdit({
         request: input.request,
         originalPrompt,
@@ -46456,6 +46490,7 @@ function createRolePlayRouter(t3) {
       const proposal = createPromptEditProposal({
         agentId: input.agentId,
         step: input.step,
+        variantId,
         originalPrompt,
         suggestedPrompt
       });
@@ -46466,7 +46501,12 @@ function createRolePlayRouter(t3) {
     acceptPromptEdit: t3.procedure.input(external_exports.object({ id: external_exports.string().uuid() })).mutation(async ({ input }) => {
       const proposal = getPromptEditProposal(input.id);
       if (!proposal) throw new Error("This prompt suggestion is no longer available");
-      await persistPromptUpdate(proposal.agentId, proposal.step, proposal.suggestedPrompt);
+      await persistPromptUpdate(
+        proposal.agentId,
+        proposal.step,
+        proposal.suggestedPrompt,
+        proposal.variantId
+      );
       deletePromptEditProposal(proposal.id);
       return { ok: true, prompt: proposal.suggestedPrompt };
     }),
@@ -46475,10 +46515,16 @@ function createRolePlayRouter(t3) {
       external_exports.object({
         agentId: AgentIdSchema,
         step: external_exports.string().trim().min(1).max(512),
-        schema: external_exports.string().max(1e6)
+        schema: external_exports.string().max(1e6),
+        variant: external_exports.string().trim().min(1).max(128).nullable().optional()
       })
     ).mutation(async ({ input }) => {
-      await persistSchemaUpdate(input.agentId, input.step, input.schema);
+      await persistSchemaUpdate(
+        input.agentId,
+        input.step,
+        input.schema,
+        input.variant ?? void 0
+      );
       bumpImportNonce();
       return {
         ok: true,
@@ -46488,10 +46534,15 @@ function createRolePlayRouter(t3) {
     userMessageHistory: t3.procedure.input(
       external_exports.object({
         agentId: AgentIdSchema,
-        region: external_exports.string().trim().min(1).max(128).nullable().optional()
+        region: external_exports.string().trim().min(1).max(128).nullable().optional(),
+        variant: external_exports.string().trim().min(1).max(128).nullable().optional()
       })
     ).query(async ({ input }) => {
-      const stateKey = resolveAgentStateKey(input.agentId, input.region ?? null);
+      const stateKey = resolveAgentStateKey(
+        input.agentId,
+        input.region ?? null,
+        input.variant ?? null
+      );
       const filePath = resolveAgentHistoryFile(stateKey);
       const states = await loadJsonArray(filePath);
       const byText = /* @__PURE__ */ new Map();
@@ -46527,18 +46578,24 @@ function createRolePlayRouter(t3) {
       external_exports.object({
         agentId: AgentIdSchema,
         region: external_exports.string().trim().min(1).max(128).nullable().optional(),
+        variant: external_exports.string().trim().min(1).max(128).nullable().optional(),
         context: external_exports.record(external_exports.string(), external_exports.unknown()).nullable().optional(),
         promptOverrides: external_exports.record(external_exports.string(), external_exports.string().max(1e6)).optional(),
         count: external_exports.number().int().min(1).max(9).default(9)
       })
     ).mutation(async ({ input }) => {
-      const stateKey = resolveAgentStateKey(input.agentId, input.region ?? null);
+      const stateKey = resolveAgentStateKey(
+        input.agentId,
+        input.region ?? null,
+        input.variant ?? null
+      );
       const rawStates = await fsStatePersistence.load(stateKey);
       const states = Array.isArray(rawStates) ? rawStates : [];
       const context = input.context ?? (await fsContextPersistence.load("default") || {});
       const variants = await generateReplyVariants({
         agentId: input.agentId,
         region: input.region ?? null,
+        variantId: input.variant ?? void 0,
         states,
         count: input.count,
         context,
@@ -46557,13 +46614,18 @@ function createRolePlayRouter(t3) {
       external_exports.object({
         agentId: AgentIdSchema,
         region: external_exports.string().trim().min(1).max(128).nullable().optional(),
+        variant: external_exports.string().trim().min(1).max(128).nullable().optional(),
         state: external_exports.object({
           history: external_exports.unknown(),
           schedule: external_exports.unknown()
         })
       })
     ).mutation(async ({ input }) => {
-      const stateKey = resolveAgentStateKey(input.agentId, input.region ?? null);
+      const stateKey = resolveAgentStateKey(
+        input.agentId,
+        input.region ?? null,
+        input.variant ?? null
+      );
       const rawStates = await fsStatePersistence.load(stateKey);
       const states = Array.isArray(rawStates) ? rawStates : [];
       const latestState = states.at(-1);
@@ -46588,8 +46650,9 @@ function createRolePlayRouter(t3) {
     })
   });
 }
-function resolveAgentStateKey(agentId, region) {
-  return region ? `${agentId}.${region}` : agentId;
+function resolveAgentStateKey(agentId, region, variant) {
+  const key = region ? `${agentId}.${region}` : agentId;
+  return variant ? `${key}.${variant}` : key;
 }
 function resolveAgentHistoryFile(agentId) {
   const baseDir = getProjectDataDir("history");
@@ -46614,23 +46677,23 @@ async function loadJsonArray(filePath) {
 }
 var noop2 = () => void 0;
 var noopAsync = () => Promise.resolve();
-async function persistPromptUpdate(agentId, stepName, prompt) {
-  const { basePath, steps } = await loadAgentSteps(agentId);
+async function persistPromptUpdate(agentId, stepName, prompt, variantId) {
+  const { basePath, steps } = await loadAgentSteps(agentId, variantId);
   const targetStep = steps.find((step) => step?.name === stepName);
   if (!targetStep || targetStep?.prompt?.type !== "file") return;
   const promptPath = path12.resolve(basePath, targetStep.prompt.path);
   await fs6.promises.writeFile(promptPath, prompt, "utf-8");
 }
-async function readEditablePrompt(agentId, stepName) {
-  const { steps } = await loadAgentSteps(agentId);
+async function readEditablePrompt(agentId, stepName, variantId) {
+  const { steps } = await loadAgentSteps(agentId, variantId);
   const targetStep = steps.find((step) => step?.name === stepName);
   if (!targetStep || targetStep?.prompt?.type !== "file") {
     throw new Error(`Step '${stepName}' does not have an editable file prompt`);
   }
   return targetStep.prompt.content();
 }
-async function persistSchemaUpdate(agentId, stepName, schema2) {
-  const { basePath, steps } = await loadAgentSteps(agentId);
+async function persistSchemaUpdate(agentId, stepName, schema2, variantId) {
+  const { basePath, steps } = await loadAgentSteps(agentId, variantId);
   const targetStep = steps.find((step) => step?.name === stepName);
   if (!targetStep || targetStep?.prompt?.type !== "file" || !targetStep.schema) {
     throw new Error(`Step '${stepName}' does not have an editable file schema`);
@@ -46640,7 +46703,7 @@ async function persistSchemaUpdate(agentId, stepName, schema2) {
   await fs6.promises.access(schemaPath);
   await validateAndReplaceSchemaFile(schemaPath, schema2);
 }
-async function loadAgentSteps(agentId) {
+async function loadAgentSteps(agentId, variantId) {
   const { factory, basePath } = await getAgentFactory(agentId);
   const ctx = createContext4({});
   const steps = [];
@@ -46652,6 +46715,7 @@ async function loadAgentSteps(agentId) {
   };
   await factory({
     ctx,
+    variantId,
     getMessages: () => [],
     logger: asLogger(console),
     scheduler: {
@@ -46882,6 +46946,7 @@ async function loadAgents(props) {
   const {
     agentId,
     region,
+    variantId,
     getMessages,
     sendMessage,
     stepTracer,
@@ -46942,6 +47007,7 @@ async function loadAgents(props) {
       stepTracer,
       stepRegistry,
       logger,
+      variantId,
       scheduler,
       getMessages: () => history,
       async sendMessage(message) {
@@ -47346,6 +47412,7 @@ var RecombineAgentProvider = class {
       await loadAgents({
         agentId: vars.agentId,
         region: vars.region,
+        variantId: vars.variantId,
         getMessages: () => conv,
         sendMessage: async (message) => {
           const m2 = typeof message === "string" ? { sender: "agent", text: message } : message;
@@ -47574,6 +47641,7 @@ async function runPromptTests(suite, selection, requestedConfig, opts) {
         agentId: suite.agentId,
         testId: String(test.id ?? ""),
         ...suite.region ? { region: suite.region } : {},
+        ...suite.variantId ? { variantId: suite.variantId } : {},
         context: test.context,
         step: test.step,
         promptOverrides: opts?.promptOverrides ?? {},
@@ -48254,53 +48322,101 @@ function envTextOrDefault2(key, defaultValue) {
 }
 
 // ../bosun-app/src/backend/tests/router.ts
-async function resolveSuiteSelection(agentId, region) {
+async function resolveSuiteSelection(agentId, region, variant) {
+  const variantId = variant || void 0;
   const bosunApi = await getProjectBosunApi();
   const meta3 = bosunApi ? await bosunApi.getAgentMeta(agentId) : null;
   const isRegionScoped = meta3?.testScope === "region";
-  if (!isRegionScoped) return { suiteKey: agentId, agentId };
+  if (!isRegionScoped) return { suiteKey: agentId, agentId, variantId };
   if (!bosunApi) {
     throw new Error(`Project does not export bosunApi, cannot resolve suite for '${agentId}'.`);
   }
-  const resolved = await bosunApi.resolveTestSuite({ agentId, region });
+  const resolved = await bosunApi.resolveTestSuite({ agentId, region, variantId });
   return {
     suiteKey: resolved.suiteKey,
     agentId,
-    region: resolved.resolvedRegion ?? void 0
+    region: resolved.resolvedRegion ?? void 0,
+    variantId
   };
 }
 function createTestsRouter(t3) {
   return t3.router({
-    list: t3.procedure.input(external_exports.object({ agent: external_exports.string(), region: external_exports.string().nullish() })).query(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+    list: t3.procedure.input(
+      external_exports.object({
+        agent: external_exports.string(),
+        region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish()
+      })
+    ).query(async ({ input }) => {
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return loadTests(suiteKey);
     }),
-    latestResult: t3.procedure.input(external_exports.object({ agent: external_exports.string(), region: external_exports.string().nullish() })).query(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+    latestResult: t3.procedure.input(
+      external_exports.object({
+        agent: external_exports.string(),
+        region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish()
+      })
+    ).query(async ({ input }) => {
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return loadLatestResultsByTest(suiteKey);
     }),
-    runProgress: t3.procedure.input(external_exports.object({ agent: external_exports.string(), region: external_exports.string().nullish() })).query(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+    runProgress: t3.procedure.input(
+      external_exports.object({
+        agent: external_exports.string(),
+        region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish()
+      })
+    ).query(async ({ input }) => {
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return getPromptTestsProgress(suiteKey);
     }),
-    historyList: t3.procedure.input(external_exports.object({ agent: external_exports.string(), region: external_exports.string().nullish() })).query(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+    historyList: t3.procedure.input(
+      external_exports.object({
+        agent: external_exports.string(),
+        region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish()
+      })
+    ).query(async ({ input }) => {
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return listRunHistory(suiteKey);
     }),
     historyRun: t3.procedure.input(
       external_exports.object({
         agent: external_exports.string(),
         region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish(),
         timestamp: external_exports.string()
       })
     ).query(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return getRunHistory(suiteKey, input.timestamp);
     }),
     resultEntry: t3.procedure.input(
       external_exports.object({
         agent: external_exports.string(),
         region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish(),
         testId: external_exports.string(),
         // TODO resultKey should be keyof of ResultEntry
         resultKey: external_exports.string(),
@@ -48308,7 +48424,11 @@ function createTestsRouter(t3) {
         timestamp: external_exports.string().optional()
       })
     ).query(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return getResultEntry(
         suiteKey,
         input.testId,
@@ -48321,12 +48441,17 @@ function createTestsRouter(t3) {
       external_exports.object({
         agent: external_exports.string(),
         region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish(),
         testId: external_exports.string(),
         index: external_exports.number().int().nonnegative(),
         timestamp: external_exports.string().optional()
       })
     ).mutation(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return analyzeRunQuotes({
         agent: suiteKey,
         testId: input.testId,
@@ -48343,28 +48468,48 @@ function createTestsRouter(t3) {
     ).mutation(async ({ input }) => {
       return findPromptIssues(input);
     }),
-    suite: t3.procedure.input(external_exports.object({ agent: external_exports.string(), region: external_exports.string().nullish() })).query(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+    suite: t3.procedure.input(
+      external_exports.object({
+        agent: external_exports.string(),
+        region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish()
+      })
+    ).query(async ({ input }) => {
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return readSuite(suiteKey);
     }),
     addTest: t3.procedure.input(
       external_exports.object({
         agent: external_exports.string(),
         region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish(),
         seed: TestRowSchema.omit({ id: true, order: true }).partial()
       })
     ).mutation(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return addTest(suiteKey, input.seed);
     }),
     createPiiSanitizedTestDraft: t3.procedure.input(
       external_exports.object({
         agent: external_exports.string(),
         region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish(),
         seed: TraceTestSeedSchema
       })
     ).mutation(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       try {
         const draft = await createPiiSanitizedTestDraft({ suiteKey, seed: input.seed });
         return { status: "ready", ...draft };
@@ -48381,11 +48526,16 @@ function createTestsRouter(t3) {
       external_exports.object({
         agent: external_exports.string(),
         region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish(),
         draftId: external_exports.string().uuid(),
         reviewed: TraceTestReviewContentSchema
       })
     ).mutation(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return createTestFromPiiSanitizedTestDraft({
         suiteKey,
         draftId: input.draftId,
@@ -48396,17 +48546,23 @@ function createTestsRouter(t3) {
       external_exports.object({
         agent: external_exports.string(),
         region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish(),
         id: external_exports.string(),
         patch: TestRowSchema.omit({ id: true, order: true }).partial()
       })
     ).mutation(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return patchTest(suiteKey, input.id, input.patch);
     }),
     patchTestsBatch: t3.procedure.input(
       external_exports.object({
         agent: external_exports.string(),
         region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish(),
         updates: external_exports.array(
           external_exports.object({
             id: external_exports.string(),
@@ -48415,21 +48571,41 @@ function createTestsRouter(t3) {
         )
       })
     ).mutation(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return patchTestsBatch(suiteKey, input.updates);
     }),
-    deleteTest: t3.procedure.input(external_exports.object({ agent: external_exports.string(), region: external_exports.string().nullish(), id: external_exports.string() })).mutation(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+    deleteTest: t3.procedure.input(
+      external_exports.object({
+        agent: external_exports.string(),
+        region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish(),
+        id: external_exports.string()
+      })
+    ).mutation(async ({ input }) => {
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return deleteTest(suiteKey, input.id);
     }),
     reorder: t3.procedure.input(
       external_exports.object({
         agent: external_exports.string(),
         region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish(),
         order: external_exports.array(external_exports.string())
       })
     ).mutation(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return reorderTests(suiteKey, input.order);
     }),
     runTests: t3.procedure.input(
@@ -48437,6 +48613,7 @@ function createTestsRouter(t3) {
         external_exports.object({
           agent: external_exports.string(),
           region: external_exports.string().nullish(),
+          variant: external_exports.string().nullish(),
           step: external_exports.string(),
           scope: external_exports.enum(["pinned", "all"]).optional(),
           promptOverrides: external_exports.record(external_exports.string(), external_exports.string().max(1e6)).optional(),
@@ -48446,6 +48623,7 @@ function createTestsRouter(t3) {
         external_exports.object({
           agent: external_exports.string(),
           region: external_exports.string().nullish(),
+          variant: external_exports.string().nullish(),
           testId: external_exports.string(),
           promptOverrides: external_exports.record(external_exports.string(), external_exports.string().max(1e6)).optional(),
           repeats: external_exports.number().int().min(1).max(500),
@@ -48453,7 +48631,11 @@ function createTestsRouter(t3) {
         }).strict()
       ])
     ).mutation(async ({ input }) => {
-      const selection = await resolveSuiteSelection(input.agent, input.region);
+      const selection = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       const runConfig = { repeats: input.repeats, concurrency: input.concurrency };
       if ("testId" in input) {
         return runPromptTestsManaged(
@@ -48473,8 +48655,18 @@ function createTestsRouter(t3) {
         input.promptOverrides
       );
     }),
-    cancelRun: t3.procedure.input(external_exports.object({ agent: external_exports.string(), region: external_exports.string().nullish() })).mutation(async ({ input }) => {
-      const { suiteKey } = await resolveSuiteSelection(input.agent, input.region);
+    cancelRun: t3.procedure.input(
+      external_exports.object({
+        agent: external_exports.string(),
+        region: external_exports.string().nullish(),
+        variant: external_exports.string().nullish()
+      })
+    ).mutation(async ({ input }) => {
+      const { suiteKey } = await resolveSuiteSelection(
+        input.agent,
+        input.region,
+        input.variant
+      );
       return cancelPromptTests(suiteKey);
     })
   });