@algolia/wizard 0.102.0 → 0.103.0-rc.217.754

package/dist/main.js

@@ -1616,6 +1616,10 @@ async function clearWorkflowState(workflowId) {
 }
 
 // src/lib/modelProfiles.ts
+var MODEL_CAPABILITIES = {
+  "claude-haiku-5-5": { forcedToolUse: true, lowestThinking: "disabled" },
+  "claude-sonnet-5-5": { forcedToolUse: false, lowestThinking: "between_tools" }
+};
 var TOKENS_PER_MILLION = 1e6;
 var INITIAL_MODEL_ATTEMPT = 1;
 var NO_TOKENS = 0;
@@ -1627,57 +1631,57 @@ var IMPLEMENTATION_MAX_OUTPUT_TOKENS = 65536;
 var IMPLEMENTATION_MAX_STEPS = 48;
 var VALIDATION_MAX_OUTPUT_TOKENS = 8192;
 var VALIDATION_MAX_STEPS = 24;
-var HAIKU_4_5_PRICES = {
-  input: 1,
-  output: 5,
-  cacheRead: 0.1,
-  cacheWrite: 1.25
+var HAIKU_5_5_PRICES = {
+  input: 0.1,
+  output: 0.5,
+  cacheRead: 0.01,
+  cacheWrite: 0.125
 };
-var SONNET_5_PRICES = {
+var SONNET_5_5_PRICES = {
   input: 2,
   output: 10,
-  cacheRead: 0.2,
+  cacheRead: 0.1,
   cacheWrite: 2.5
 };
 var MODEL_PROFILES = {
   quick: {
-    model: "claude-haiku-4-5",
+    model: "claude-haiku-5-5",
     thinking: "disabled",
     maxOutputTokens: QUICK_MAX_OUTPUT_TOKENS,
     maxSteps: QUICK_MAX_STEPS,
-    prices: HAIKU_4_5_PRICES
+    prices: HAIKU_5_5_PRICES
   },
   analysis: {
-    model: "claude-sonnet-5",
+    model: "claude-sonnet-5-5",
     thinking: "adaptive",
     effort: "low",
     maxOutputTokens: ANALYSIS_MAX_OUTPUT_TOKENS,
     maxSteps: ANALYSIS_MAX_STEPS,
-    prices: SONNET_5_PRICES
+    prices: SONNET_5_5_PRICES
   },
   implementation: {
-    model: "claude-sonnet-5",
+    model: "claude-sonnet-5-5",
     thinking: "adaptive",
     effort: "medium",
     maxOutputTokens: IMPLEMENTATION_MAX_OUTPUT_TOKENS,
     maxSteps: IMPLEMENTATION_MAX_STEPS,
-    prices: SONNET_5_PRICES
+    prices: SONNET_5_5_PRICES
   },
   implementationRetry: {
-    model: "claude-sonnet-5",
+    model: "claude-sonnet-5-5",
     thinking: "adaptive",
     effort: "high",
     maxOutputTokens: IMPLEMENTATION_MAX_OUTPUT_TOKENS,
     maxSteps: IMPLEMENTATION_MAX_STEPS,
-    prices: SONNET_5_PRICES
+    prices: SONNET_5_5_PRICES
   },
   validation: {
-    model: "claude-sonnet-5",
+    model: "claude-sonnet-5-5",
     thinking: "adaptive",
     effort: "low",
     maxOutputTokens: VALIDATION_MAX_OUTPUT_TOKENS,
     maxSteps: VALIDATION_MAX_STEPS,
-    prices: SONNET_5_PRICES
+    prices: SONNET_5_5_PRICES
   }
 };
 function getModelProfile(profileName) {
@@ -1690,6 +1694,14 @@ function providerOptionsForProfile(profileName) {
     ...profile.effort !== void 0 && { effort: profile.effort }
   };
 }
+function toolChoiceForProfile(profileName) {
+  const { model } = getModelProfile(profileName);
+  return MODEL_CAPABILITIES[model].forcedToolUse ? "required" : "auto";
+}
+function lowestThinkingForProfile(profileName) {
+  const { model } = getModelProfile(profileName);
+  return { type: MODEL_CAPABILITIES[model].lowestThinking };
+}
 function tokenCount(value = NO_TOKENS) {
   return value;
 }
@@ -2447,7 +2459,7 @@ async function ensureApplication(resumeAppId) {
 // package.json
 var package_default = {
   name: "@algolia/wizard",
-  version: "0.102.0",
+  version: "0.103.0-rc.217.754",
   description: "Magically implement Algolia functionality in your codebase",
   type: "module",
   engines: {
@@ -2494,11 +2506,11 @@ var package_default = {
     vitest: "^4.1.11"
   },
   dependencies: {
-    "@ai-sdk/anthropic": "^3.0.118",
-    "@ai-sdk/openai-compatible": "^2.0.75",
+    "@ai-sdk/anthropic": "^3.0.128",
+    "@ai-sdk/openai-compatible": "^2.0.81",
     "@hono/node-server": "^2.0.10",
     "@segment/analytics-node": "^3.1.0",
-    ai: "^6.0.283",
+    ai: "^6.0.302",
     "cross-keychain": "^1.1.0",
     dotenv: "^17.4.2",
     hono: "^4.13.8",
@@ -5418,6 +5430,10 @@ var MISSING_REPORT_STATUS_ERROR_MESSAGE = "Agent finished without calling report
 var MISSING_REPORT_USER_MESSAGE = "This step ran into a problem finishing. Run the wizard again to retry it.";
 var REPORT_STATUS_RETRIES = 2;
 var STEP_LIMIT_INSTRUCTION = "When reportStatus is the only tool available, you reached the step limit. This limit is not a failure, so do not report fail because of it: report status=success with the work done so far, and name the remaining work in the summary.";
+var TOOL_EVERY_TURN_INSTRUCTION = "Call at least one tool in every response. Do not reply with text only. The run ends only when you call reportStatus.";
+var ENDED_TURN_NUDGE = "Your last response did not call a tool. Continue the work with tools. When the work is done, call reportStatus.";
+var INITIAL_PROMPT = "Follow system instructions";
+var ENDED_TURN_CONTINUATIONS = 3;
 var RECENT_STEPS_LOGGED = 3;
 var LOGGED_CHARS = 300;
 var ATTEMPT_NUMBER_OFFSET = 1;
@@ -5455,6 +5471,13 @@ var MissingReportError = class extends Error {
     this.incompleteCause = incompleteCause;
   }
 };
+var EndedTurnError = class extends MissingReportError {
+  continuation;
+  constructor(continuation) {
+    super(null);
+    this.continuation = continuation;
+  }
+};
 function incompleteCauseOf(stepCount, maxSteps, finishReason) {
   if (stepCount >= maxSteps) return "stepLimit";
   if (finishReason === "length") return "truncated";
@@ -5513,10 +5536,23 @@ function exhaustedError(kind, err) {
 }
 async function runAgent(req) {
   let retryReason = null;
-  for (let attempt = INITIAL_ATTEMPT; attempt <= REPORT_STATUS_RETRIES; attempt++) {
+  let attempt = INITIAL_ATTEMPT;
+  let continuation;
+  let continuations = 0;
+  for (; ; ) {
     try {
-      return await runAgentAttempt(req, attempt, retryReason);
+      return await runAgentAttempt(req, attempt, retryReason, continuation);
     } catch (err) {
+      continuation = void 0;
+      if (err instanceof EndedTurnError && continuations < ENDED_TURN_CONTINUATIONS) {
+        continuations++;
+        continuation = err.continuation;
+        logger.warn(
+          { attempt: attempt + 1, continuations },
+          "continuing runAgent after a turn ended without a tool call"
+        );
+        continue;
+      }
       const kind = retryKind(err);
       if (!kind) throw err;
       if (attempt === REPORT_STATUS_RETRIES) throw exhaustedError(kind, err);
@@ -5525,19 +5561,23 @@ async function runAgent(req) {
         { attempt: attempt + 1, err },
         isMissingReportRetry(kind) ? "retrying runAgent after missing reportStatus" : "retrying runAgent after a transient provider error"
       );
+      attempt++;
     }
   }
-  throw new Error("unreachable");
 }
-async function runAgentAttempt(req, attempt, retryReason) {
+async function runAgentAttempt(req, attempt, retryReason, continuation) {
   const start = Date.now();
   const profileName = req.modelProfile ?? "implementation" /* implementation */;
   const profile = getModelProfile(profileName);
   const profileOptions = providerOptionsForProfile(profileName);
+  const toolChoice = toolChoiceForProfile(profileName);
   const maxSteps = req.maxSteps ?? profile.maxSteps;
+  const stepsUsed = continuation?.stepsUsed ?? 0;
+  const stepsLeft = maxSteps - stepsUsed;
   const forceReport = retryReason === "missingReport";
   const incompleteReason = incompleteReasonOf(retryReason);
-  const modelOptions = forceReport ? { thinking: { type: "disabled" } } : profileOptions;
+  const reportThinking = lowestThinkingForProfile(profileName);
+  const modelOptions = forceReport ? { thinking: reportThinking, effort: void 0 } : profileOptions;
   logger.info(
     {
       startedAt: new Date(start).toISOString(),
@@ -5545,8 +5585,9 @@ async function runAgentAttempt(req, attempt, retryReason) {
       profile: profileName,
       model: profile.model,
       effort: forceReport ? void 0 : profile.effort,
-      thinking: forceReport ? "disabled" : profile.thinking,
-      retryReason
+      thinking: forceReport ? reportThinking.type : profile.thinking,
+      retryReason,
+      stepsUsed
     },
     "runAgent started"
   );
@@ -5560,7 +5601,7 @@ async function runAgentAttempt(req, attempt, retryReason) {
     fetch: proxyFetch
   });
   const toolContext = req.toolContext ?? createToolContext();
-  if (attempt > INITIAL_ATTEMPT) resetToolCounts(toolContext);
+  if (attempt > INITIAL_ATTEMPT && !continuation) resetToolCounts(toolContext);
   const readTools = ["readFile", "searchFiles", "listFiles"];
   const hasReadTools = !req.tools || req.tools.some((t) => readTools.includes(t));
   const instructions = [
@@ -5569,6 +5610,7 @@ async function runAgentAttempt(req, attempt, retryReason) {
       "When you need to read or search multiple files, issue those tool calls together in one step rather than one at a time."
     ] : [],
     "Call notifyUser whenever you start a new phase of work or your focus shifts (e.g. moving from reading code to writing files) \u2014 a short, plain-language, high-level update. Do not call it for every tool use.",
+    ...toolChoice === "auto" ? [TOOL_EVERY_TURN_INSTRUCTION] : [],
     STEP_LIMIT_INSTRUCTION,
     // Last so a retry does not bust the cached tools+system prefix from
     // the first attempt. Inspect-and-continue, not start over: implement
@@ -5613,11 +5655,11 @@ async function runAgentAttempt(req, attempt, retryReason) {
       output: req.outputSchema,
       tools: req.tools
     }),
-    toolChoice: "required",
-    stopWhen: [reportAccepted, stepCountIs(maxSteps)],
+    toolChoice,
+    stopWhen: [reportAccepted, stepCountIs(stepsLeft)],
     prepareStep: ({ stepNumber, messages }) => ({
       messages: withCachedHistory(messages),
-      ...stepNumber >= maxSteps - 1 && {
+      ...stepNumber >= stepsLeft - 1 && {
         activeTools: ["reportStatus"]
       }
     }),
@@ -5625,9 +5667,11 @@ async function runAgentAttempt(req, attempt, retryReason) {
     ...{ onError: () => {
     } }
   });
-  const stream = await agent.stream({
-    prompt: "Follow system instructions"
-  });
+  const prompt = continuation ? [
+    ...continuation.messages,
+    { role: "user", content: ENDED_TURN_NUDGE }
+  ] : [{ role: "user", content: INITIAL_PROMPT }];
+  const stream = await agent.stream({ prompt });
   let chunks = [];
   const chunkLimit = 5;
   let streamError;
@@ -5659,7 +5703,8 @@ async function runAgentAttempt(req, attempt, retryReason) {
   const usage = await stream.totalUsage;
   const steps = await stream.steps;
   const stepCount = steps.length;
-  const hitStepLimit = stepCount >= maxSteps;
+  const totalSteps = stepsUsed + stepCount;
+  const hitStepLimit = totalSteps >= maxSteps;
   const lastStep = steps[steps.length - 1];
   const finishReason = lastStep?.finishReason;
   trackAgentRun({
@@ -5707,8 +5752,16 @@ async function runAgentAttempt(req, attempt, retryReason) {
       },
       "Agent finished without calling reportStatus"
     );
+    const endedTurn = toolChoice === "auto" && finishReason === "stop" && !lastStep?.toolCalls?.length && !hitStepLimit;
+    if (endedTurn) {
+      const { messages } = await stream.response;
+      throw new EndedTurnError({
+        messages: [...prompt, ...messages],
+        stepsUsed: totalSteps
+      });
+    }
     throw new MissingReportError(
-      incompleteCauseOf(stepCount, maxSteps, finishReason)
+      incompleteCauseOf(totalSteps, maxSteps, finishReason)
     );
   }
   const result = report.output;
@@ -10073,7 +10126,7 @@ async function resetProjectState() {
 // package.json with { type: 'json' }
 var package_default2 = {
   name: "@algolia/wizard",
-  version: "0.102.0",
+  version: "0.103.0-rc.217.754",
   description: "Magically implement Algolia functionality in your codebase",
   type: "module",
   engines: {
@@ -10120,11 +10173,11 @@ var package_default2 = {
     vitest: "^4.1.11"
   },
   dependencies: {
-    "@ai-sdk/anthropic": "^3.0.118",
-    "@ai-sdk/openai-compatible": "^2.0.75",
+    "@ai-sdk/anthropic": "^3.0.128",
+    "@ai-sdk/openai-compatible": "^2.0.81",
     "@hono/node-server": "^2.0.10",
     "@segment/analytics-node": "^3.1.0",
-    ai: "^6.0.283",
+    ai: "^6.0.302",
     "cross-keychain": "^1.1.0",
     dotenv: "^17.4.2",
     hono: "^4.13.8",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@algolia/wizard",
-  "version": "0.102.0",
+  "version": "0.103.0-rc.217.754",
   "description": "Magically implement Algolia functionality in your codebase",
   "type": "module",
   "engines": {
@@ -47,11 +47,11 @@
     "vitest": "^4.1.11"
   },
   "dependencies": {
-    "@ai-sdk/anthropic": "^3.0.118",
-    "@ai-sdk/openai-compatible": "^2.0.75",
+    "@ai-sdk/anthropic": "^3.0.128",
+    "@ai-sdk/openai-compatible": "^2.0.81",
     "@hono/node-server": "^2.0.10",
     "@segment/analytics-node": "^3.1.0",
-    "ai": "^6.0.283",
+    "ai": "^6.0.302",
     "cross-keychain": "^1.1.0",
     "dotenv": "^17.4.2",
     "hono": "^4.13.8",