@kody-ade/kody-engine 0.4.592 → 0.4.593

package/dist/bin/kody.js

@@ -15,7 +15,7 @@ var init_package = __esm({
   "package.json"() {
     package_default = {
       name: "@kody-ade/kody-engine",
-      version: "0.4.592",
+      version: "0.4.593",
       description: "kody \u2014 autonomous development engine. Single-session Claude Code agent behind a generic executor + declarative implementation profiles.",
       license: "MIT",
       type: "module",
@@ -234,6 +234,32 @@ function parseModelRuntimeConfig(modelSpec, rawConfig) {
   if (protocol === "openai") out.litellmProvider = "openai";
   return out;
 }
+function parseAutomaticModels(raw) {
+  if (raw === void 0) return void 0;
+  if (!Array.isArray(raw)) throw new Error("kody.config.json: agent.automaticModels must be an array");
+  const models = raw.map((entry, index) => {
+    if (!entry || typeof entry !== "object" || Array.isArray(entry)) {
+      throw new Error(`kody.config.json: agent.automaticModels[${index}] must be an object`);
+    }
+    const record2 = entry;
+    const spec = optionalRuntimeString(record2, "spec");
+    const modelName = optionalRuntimeString(record2, "modelName");
+    const provider = optionalRuntimeString(record2, "provider");
+    if (!spec || !modelName || !provider) {
+      throw new Error(`kody.config.json: agent.automaticModels[${index}] requires spec, provider, and modelName`);
+    }
+    const model = { provider, model: modelName, spec };
+    const protocol = optionalRuntimeString(record2, "protocol");
+    const baseURL = optionalRuntimeString(record2, "baseURL");
+    const apiKeyEnvVar = optionalRuntimeString(record2, "apiKeyEnvVar");
+    if (protocol) model.protocol = protocol;
+    if (baseURL) model.baseURL = baseURL;
+    if (apiKeyEnvVar) model.apiKeyEnvVar = apiKeyEnvVar;
+    if (protocol === "openai") model.litellmProvider = "openai";
+    return model;
+  });
+  return models.length > 0 ? models : void 0;
+}
 function litellmModelGroup(model) {
   return model.spec?.trim() || model.model;
 }
@@ -267,6 +293,10 @@ function loadConfig(projectDir = process.cwd()) {
   if (!github.owner || !github.repo) {
     throw new Error(`kody.config.json: github.owner and github.repo are required`);
   }
+  const automaticModels = parseAutomaticModels(agent.automaticModels);
+  if (agent.model === "automatic" && (!automaticModels || automaticModels.length < 2)) {
+    throw new Error("kody.config.json: Automatic requires at least two agent.automaticModels");
+  }
   return {
     quality: {
       typecheck: typeof quality.typecheck === "string" ? quality.typecheck : "",
@@ -284,6 +314,7 @@ function loadConfig(projectDir = process.cwd()) {
     },
     agent: {
       model: String(agent.model),
+      ...automaticModels ? { automaticModels } : {},
       ...parsePerImplementation(agent.perImplementation),
       ...parsePerImplementationReasoningEffort(agent.perImplementationReasoningEffort),
       ...parseAgentReasoningEffort(agent.reasoningEffort)
@@ -4169,6 +4200,7 @@ async function runAgent(opts) {
   const missingParentWriteGuard = createMissingParentWriteGuard(opts.cwd);
   const outputContractPostWriteHook = opts.outputContract ? createOutputContractPostWriteHook(opts.outputContract) : null;
   const outputContractStopHook = opts.outputContract ? createOutputContractStopHook(opts.outputContract) : null;
+  let finalSafeToReplay = true;
   for (let attempt = 0; ; attempt++) {
     let ndjsonWriteFailed = false;
     let ndjsonWriteError;
@@ -4496,7 +4528,8 @@ async function runAgent(opts) {
         errorMessage2 = e instanceof Error ? e.message : String(e);
       } else {
         outcome = "failed";
-        outcomeKind = "model_error";
+        const message = e instanceof Error ? e.message : String(e);
+        outcomeKind = /\b429\b|rate[ _-]?limit|too many requests/i.test(message) ? "rate_limit" : "model_error";
         errorMessage2 = e instanceof Error ? e.message : String(e);
       }
     } finally {
@@ -4528,6 +4561,7 @@ async function runAgent(opts) {
       }
     }
     const shouldRetry = outcome === "failed" && attempt < MAX_CONNECTION_RETRIES && !sawMutatingTool && (isTransientConnectionError(errorMessage2) || noWorkSuccess);
+    finalSafeToReplay = !sawMutatingTool;
     if (!shouldRetry) break;
     const delayMs = CONNECTION_RETRY_BASE_MS * 2 ** attempt;
     process.stderr.write(
@@ -4548,6 +4582,7 @@ async function runAgent(opts) {
   return {
     outcome,
     outcomeKind,
+    safeToReplay: finalSafeToReplay,
     finalText,
     ...submittedState ? { submittedState } : {},
     error: errorMessage2,
@@ -22286,17 +22321,18 @@ async function runImplementation(profileName, input) {
   const modelSpec = perImplementationModel ? perImplementationModel : profile.claudeCode.model === "inherit" ? config.agent.model : profile.claudeCode.model;
   const profileHasThinkingTokens = typeof profile.claudeCode.maxThinkingTokens === "number" && profile.claudeCode.maxThinkingTokens > 0;
   const reasoningEffort = config.agent.perImplementationReasoningEffort?.[profileName] ?? profile.claudeCode.reasoningEffort ?? (profileHasThinkingTokens ? void 0 : config.agent.reasoningEffort);
-  let model;
+  let modelCandidates;
   try {
-    model = parseProviderModel(modelSpec);
+    modelCandidates = modelSpec === "automatic" ? config.agent.automaticModels ?? [] : [parseProviderModel(modelSpec)];
+    if (modelCandidates.length === 0) throw new Error("Automatic has no configured models");
   } catch (err) {
     return finishAndEnd({
       exitCode: 99,
       reason: `agent.model invalid: ${err instanceof Error ? err.message : String(err)}`
     });
   }
-  let litellm;
-  let runtimeModelEnvironment;
+  const model = modelCandidates[0];
+  const activeLiteLLM = /* @__PURE__ */ new Set();
   const ctx = {
     args,
     cwd: input.cwd,
@@ -22393,90 +22429,107 @@ async function runImplementation(profileName, input) {
     const syntheticPath = ctx.data.syntheticPluginPath;
     const pluginPaths = [...externalPlugins, ...syntheticPath ? [syntheticPath] : []];
     const agents = loadSubagents(profile);
-    if (runtimeModelEnvironment === void 0) {
-      runtimeModelEnvironment = {};
-      if (needsLitellmProxy(model)) {
-        const resolved2 = await resolveRuntimeModelEnvironment(model, ctx);
+    let finalResult;
+    for (let candidateIndex = 0; candidateIndex < modelCandidates.length; candidateIndex++) {
+      const candidate = modelCandidates[candidateIndex];
+      let runtimeModelEnvironment = {};
+      if (needsLitellmProxy(candidate)) {
+        const resolved2 = await resolveRuntimeModelEnvironment(candidate, ctx);
         runtimeModelEnvironment = resolved2.environment;
         for (const warning of resolved2.warnings) process.stderr.write(`\u26A0 ${warning}
 `);
       }
-    }
-    if (litellm === void 0) {
+      let lm;
       try {
-        litellm = await startLitellmIfNeeded(model, input.cwd, void 0, runtimeModelEnvironment);
+        lm = await startLitellmIfNeeded(candidate, input.cwd, void 0, runtimeModelEnvironment);
+        if (lm) activeLiteLLM.add(lm);
       } catch (err) {
         throw new Error(`litellm startup failed: ${err instanceof Error ? err.message : String(err)}`);
       }
+      ctx.data.jobModelProvider = candidate.provider;
+      ctx.data.jobModelName = candidate.model;
+      const result = await runAgent({
+        prompt,
+        model: candidate,
+        cwd: input.cwd,
+        environment: {
+          ...ctx.data.capabilityEnvironment && typeof ctx.data.capabilityEnvironment === "object" && !Array.isArray(ctx.data.capabilityEnvironment) ? ctx.data.capabilityEnvironment : {},
+          ...runtimeModelEnvironment
+        },
+        litellmUrl: lm?.url ?? null,
+        // On a connection drop mid-run, restart the (possibly crashed) proxy
+        // before the agent retries. No-op for direct-Anthropic runs (lm null).
+        ensureBackend: lm ? () => lm.ensureHealthy().then(() => void 0) : void 0,
+        // Pure liveness probe so the agent can spot a hollow "success" (proxy
+        // crashed mid-request, SDK still reported success). No-op when lm null.
+        isBackendHealthy: lm ? () => lm.isHealthy() : void 0,
+        verbose: input.verbose,
+        quiet: input.quiet,
+        abortController: input.abortController,
+        deadlineAtMs: input.deadlineAtMs,
+        ndjsonDir,
+        additionalDirectories: agentTaskArtifacts ? [agentTaskArtifacts.absDir] : void 0,
+        allowedToolsOverride: profile.claudeCode.tools,
+        disallowedToolsOverride: profile.claudeCode.disallowedTools,
+        permissionModeOverride: profile.claudeCode.permissionMode,
+        mcpServers: profile.claudeCode.mcpServers.length > 0 ? profile.claudeCode.mcpServers : void 0,
+        pluginPaths: pluginPaths.length > 0 ? pluginPaths : void 0,
+        agents,
+        maxTurns: profile.claudeCode.maxTurns,
+        reasoningEffort,
+        maxThinkingTokens: profile.claudeCode.maxThinkingTokens,
+        maxTurnTimeoutMs: typeof profile.claudeCode.maxTurnTimeoutSec === "number" ? Math.floor(profile.claudeCode.maxTurnTimeoutSec * 1e3) : void 0,
+        // DISCIPLINE leads so the stable, role-agnostic block sits at the front
+        // of the cacheable system-prompt prefix; profile/task appends follow.
+        systemPromptAppend: [
+          DISCIPLINE,
+          agentIdentityBlock,
+          jobRefBlock,
+          jobWhyBlock,
+          profile.claudeCode.systemPromptAppend,
+          agentTaskArtifacts?.promptAddendum
+        ].filter((s) => typeof s === "string" && s.length > 0).join("\n\n") || void 0,
+        cacheable: profile.claudeCode.cacheable,
+        enableVerifyTool: profile.claudeCode.enableVerifyTool,
+        enableSubmitTool: profile.claudeCode.enableSubmitTool,
+        // Locked-toolbox capability mode: `loadJobFromFile` flips `ctx.data.capabilityTools`
+        // when a capability declares `tools` in profile.json. The executor doesn't need
+        // to know the palette — it just forwards the flag so agent.ts can spin
+        // up the in-process `kody-capability` MCP server with the right context.
+        enableCapabilityTool: Array.isArray(ctx.data.capabilityTools) && ctx.data.capabilityTools.length > 0,
+        capabilityOperatorMention: typeof ctx.data.capabilityOperatorMention === "string" ? ctx.data.capabilityOperatorMention : void 0,
+        // Stamp the running capability's slug onto recommendations so the dashboard
+        // keys trust per capability (not per agent). `jobSlug` is set by loadJobFromFile.
+        capabilitySlug: typeof ctx.data.jobSlug === "string" ? ctx.data.jobSlug : void 0,
+        capabilityDefaultBranch: config.git.defaultBranch,
+        // owner/repo from kody.config.json; envelope falls back to GITHUB_REPOSITORY
+        // for tester repos that don't set config.github (the file isn't always
+        // checked in). Either way, capabilityMcp needs "owner/name" to hit the compare API.
+        capabilityRepoSlug: config.github?.owner && config.github?.repo ? `${config.github.owner}/${config.github.repo}` : process.env.GITHUB_REPOSITORY?.trim() || void 0,
+        verifyToolMaxAttempts: profile.claudeCode.verifyAttempts ?? null,
+        verifyConfig: profile.claudeCode.enableVerifyTool ? config : void 0,
+        implementationName: profileName,
+        settingSources: profile.claudeCode.settingSources,
+        outputContract: typeof ctx.data.capabilityOutputPath === "string" && ctx.data.capabilityOutputSchema && typeof ctx.data.capabilityOutputSchema === "object" && !Array.isArray(ctx.data.capabilityOutputSchema) ? {
+          path: ctx.data.capabilityOutputPath,
+          schema: ctx.data.capabilityOutputSchema
+        } : void 0
+      });
+      finalResult = result;
+      const next = modelCandidates[candidateIndex + 1];
+      if (!next || result.outcome !== "failed" || result.outcomeKind !== "rate_limit" || result.safeToReplay !== true) {
+        return result;
+      }
+      process.stderr.write(
+        `[kody agent] ${candidate.spec ?? candidate.model} is rate limited; continuing with ${next.spec ?? next.model}
+`
+      );
+      if (lm) {
+        lm.kill();
+        activeLiteLLM.delete(lm);
+      }
     }
-    const lm = litellm;
-    return runAgent({
-      prompt,
-      model,
-      cwd: input.cwd,
-      environment: {
-        ...ctx.data.capabilityEnvironment && typeof ctx.data.capabilityEnvironment === "object" && !Array.isArray(ctx.data.capabilityEnvironment) ? ctx.data.capabilityEnvironment : {},
-        ...runtimeModelEnvironment
-      },
-      litellmUrl: lm?.url ?? null,
-      // On a connection drop mid-run, restart the (possibly crashed) proxy
-      // before the agent retries. No-op for direct-Anthropic runs (lm null).
-      ensureBackend: lm ? () => lm.ensureHealthy().then(() => void 0) : void 0,
-      // Pure liveness probe so the agent can spot a hollow "success" (proxy
-      // crashed mid-request, SDK still reported success). No-op when lm null.
-      isBackendHealthy: lm ? () => lm.isHealthy() : void 0,
-      verbose: input.verbose,
-      quiet: input.quiet,
-      abortController: input.abortController,
-      deadlineAtMs: input.deadlineAtMs,
-      ndjsonDir,
-      additionalDirectories: agentTaskArtifacts ? [agentTaskArtifacts.absDir] : void 0,
-      allowedToolsOverride: profile.claudeCode.tools,
-      disallowedToolsOverride: profile.claudeCode.disallowedTools,
-      permissionModeOverride: profile.claudeCode.permissionMode,
-      mcpServers: profile.claudeCode.mcpServers.length > 0 ? profile.claudeCode.mcpServers : void 0,
-      pluginPaths: pluginPaths.length > 0 ? pluginPaths : void 0,
-      agents,
-      maxTurns: profile.claudeCode.maxTurns,
-      reasoningEffort,
-      maxThinkingTokens: profile.claudeCode.maxThinkingTokens,
-      maxTurnTimeoutMs: typeof profile.claudeCode.maxTurnTimeoutSec === "number" ? Math.floor(profile.claudeCode.maxTurnTimeoutSec * 1e3) : void 0,
-      // DISCIPLINE leads so the stable, role-agnostic block sits at the front
-      // of the cacheable system-prompt prefix; profile/task appends follow.
-      systemPromptAppend: [
-        DISCIPLINE,
-        agentIdentityBlock,
-        jobRefBlock,
-        jobWhyBlock,
-        profile.claudeCode.systemPromptAppend,
-        agentTaskArtifacts?.promptAddendum
-      ].filter((s) => typeof s === "string" && s.length > 0).join("\n\n") || void 0,
-      cacheable: profile.claudeCode.cacheable,
-      enableVerifyTool: profile.claudeCode.enableVerifyTool,
-      enableSubmitTool: profile.claudeCode.enableSubmitTool,
-      // Locked-toolbox capability mode: `loadJobFromFile` flips `ctx.data.capabilityTools`
-      // when a capability declares `tools` in profile.json. The executor doesn't need
-      // to know the palette — it just forwards the flag so agent.ts can spin
-      // up the in-process `kody-capability` MCP server with the right context.
-      enableCapabilityTool: Array.isArray(ctx.data.capabilityTools) && ctx.data.capabilityTools.length > 0,
-      capabilityOperatorMention: typeof ctx.data.capabilityOperatorMention === "string" ? ctx.data.capabilityOperatorMention : void 0,
-      // Stamp the running capability's slug onto recommendations so the dashboard
-      // keys trust per capability (not per agent). `jobSlug` is set by loadJobFromFile.
-      capabilitySlug: typeof ctx.data.jobSlug === "string" ? ctx.data.jobSlug : void 0,
-      capabilityDefaultBranch: config.git.defaultBranch,
-      // owner/repo from kody.config.json; envelope falls back to GITHUB_REPOSITORY
-      // for tester repos that don't set config.github (the file isn't always
-      // checked in). Either way, capabilityMcp needs "owner/name" to hit the compare API.
-      capabilityRepoSlug: config.github?.owner && config.github?.repo ? `${config.github.owner}/${config.github.repo}` : process.env.GITHUB_REPOSITORY?.trim() || void 0,
-      verifyToolMaxAttempts: profile.claudeCode.verifyAttempts ?? null,
-      verifyConfig: profile.claudeCode.enableVerifyTool ? config : void 0,
-      implementationName: profileName,
-      settingSources: profile.claudeCode.settingSources,
-      outputContract: typeof ctx.data.capabilityOutputPath === "string" && ctx.data.capabilityOutputSchema && typeof ctx.data.capabilityOutputSchema === "object" && !Array.isArray(ctx.data.capabilityOutputSchema) ? {
-        path: ctx.data.capabilityOutputPath,
-        schema: ctx.data.capabilityOutputSchema
-      } : void 0
-    });
+    return finalResult;
   };
   ctx.data.__invokeAgent = invokeAgent;
   const cc = profile.claudeCode;
@@ -22708,7 +22761,7 @@ async function runImplementation(profileName, input) {
       }
     }
     try {
-      litellm?.kill();
+      for (const proxy of activeLiteLLM) proxy.kill();
     } catch {
     }
   }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kody-ade/kody-engine",
-  "version": "0.4.592",
+  "version": "0.4.593",
   "description": "kody — autonomous development engine. Single-session Claude Code agent behind a generic executor + declarative implementation profiles.",
   "license": "MIT",
   "type": "module",
@@ -12,30 +12,6 @@
     "templates",
     "kody.config.schema.json"
   ],
-  "scripts": {
-    "kody:run": "tsx bin/kody.ts",
-    "serve": "tsx bin/kody.ts serve",
-    "serve:vscode": "tsx bin/kody.ts serve vscode",
-    "serve:claude": "tsx bin/kody.ts serve claude",
-    "clean:dist": "node scripts/clean-dist.cjs",
-    "build": "pnpm clean:dist && tsup && node scripts/copy-assets.cjs",
-    "check:modularity": "tsx scripts/check-script-modularity.ts",
-    "pretest": "pnpm check:modularity",
-    "test": "vitest run tests/unit tests/int --coverage",
-    "posttest": "tsx scripts/check-coverage-floor.ts",
-    "test:smoke": "vitest run tests/smoke --no-coverage",
-    "test:e2e": "vitest run tests/e2e --no-coverage",
-    "verify:live-release": "tsx scripts/live-release-gate.ts",
-    "test:runtime-services": "node --test \"tests/runtime-services/*.test.mjs\"",
-    "test:all": "vitest run tests --no-coverage",
-    "typecheck": "tsc --noEmit",
-    "lint": "biome check",
-    "lint:fix": "biome check --write",
-    "format": "biome format --write",
-    "verify:package": "node scripts/verify-package-tarball.cjs",
-    "brain:publish": "docker buildx build --platform linux/amd64 -f runner/Dockerfile.brain --build-arg KODY_ENGINE_REF=$(git rev-parse HEAD) -t ghcr.io/${KODY_BRAIN_GHCR_OWNER:-aharonyaircohen}/kody-brain:latest --push runner",
-    "prepublishOnly": "pnpm typecheck && pnpm test:runtime-services && pnpm build && pnpm verify:package"
-  },
   "dependencies": {
     "@actions/cache": "^6.0.0",
     "@anthropic-ai/claude-agent-sdk": "0.2.119",
@@ -62,5 +38,28 @@
     "url": "git+https://github.com/aharonyaircohen/kody-engine.git"
   },
   "homepage": "https://github.com/aharonyaircohen/kody-engine",
-  "bugs": "https://github.com/aharonyaircohen/kody-engine/issues"
-}
+  "bugs": "https://github.com/aharonyaircohen/kody-engine/issues",
+  "scripts": {
+    "kody:run": "tsx bin/kody.ts",
+    "serve": "tsx bin/kody.ts serve",
+    "serve:vscode": "tsx bin/kody.ts serve vscode",
+    "serve:claude": "tsx bin/kody.ts serve claude",
+    "clean:dist": "node scripts/clean-dist.cjs",
+    "build": "pnpm clean:dist && tsup && node scripts/copy-assets.cjs",
+    "check:modularity": "tsx scripts/check-script-modularity.ts",
+    "pretest": "pnpm check:modularity",
+    "test": "vitest run tests/unit tests/int --coverage",
+    "posttest": "tsx scripts/check-coverage-floor.ts",
+    "test:smoke": "vitest run tests/smoke --no-coverage",
+    "test:e2e": "vitest run tests/e2e --no-coverage",
+    "verify:live-release": "tsx scripts/live-release-gate.ts",
+    "test:runtime-services": "node --test \"tests/runtime-services/*.test.mjs\"",
+    "test:all": "vitest run tests --no-coverage",
+    "typecheck": "tsc --noEmit",
+    "lint": "biome check",
+    "lint:fix": "biome check --write",
+    "format": "biome format --write",
+    "verify:package": "node scripts/verify-package-tarball.cjs",
+    "brain:publish": "docker buildx build --platform linux/amd64 -f runner/Dockerfile.brain --build-arg KODY_ENGINE_REF=$(git rev-parse HEAD) -t ghcr.io/${KODY_BRAIN_GHCR_OWNER:-aharonyaircohen}/kody-brain:latest --push runner"
+  }
+}