Address review feedback: use effort params, fix model names, add safety checks

agents/claude.ts

@@ -5,12 +5,12 @@ import { log } from "../utils/cli.ts";
 import { addInstructions } from "./instructions.ts";
 import { agent, createAgentEnv, installFromNpmTarball } from "./shared.ts";
 
-// model configuration based on effort level
-// uses model family aliases that auto-resolve to latest version
-const claudeModels: Record<Effort, { model: string; thinking: boolean }> = {
-  nothink: { model: "claude-haiku-4-5", thinking: false },
-  think: { model: "claude-sonnet-4-5", thinking: true },
-  max: { model: "claude-opus-4-5", thinking: true },
+// effort levels for Claude Code
+// Claude Code automatically selects the best model based on effort
+const claudeEffortLevels: Record<Effort, "low" | "medium" | "high"> = {
+  nothink: "low",
+  think: "medium",
+  max: "high",
 } as const;
 
 export const claude = agent({
@@ -30,9 +30,9 @@ export const claude = agent({
     const prompt = addInstructions({ payload, repo });
     log.group("Full prompt", () => log.info(prompt));
 
-    // get model configuration based on effort
-    const modelConfig = claudeModels[effort];
-    log.info(`Using model: ${modelConfig.model}, thinking: ${modelConfig.thinking}`);
+    // get effort level - Claude Code automatically selects the best model
+    const effortLevel = claudeEffortLevels[effort];
+    log.info(`Using effort: ${effortLevel}`);
 
     // SECURITY: For PUBLIC repos, Claude Code spawns subprocesses with full process.env, leaking API keys.
     // disable native Bash; agents use MCP bash tool which filters secrets.
@@ -62,15 +62,10 @@ export const claude = agent({
     const queryOptions: Options = {
       ...sandboxOptions,
       mcpServers,
-      model: modelConfig.model,
+      effort: effortLevel,
       pathToClaudeCodeExecutable: cliPath,
       env: createAgentEnv({ ANTHROPIC_API_KEY: apiKey }),
     };
-    // only set maxThinkingTokens when we want to disable thinking (0)
-    // omit the property to use default when thinking is enabled
-    if (!modelConfig.thinking) {
-      queryOptions.maxThinkingTokens = 0;
-    }
 
     const queryInstance = query({
       prompt,

agents/codex.ts

@@ -7,12 +7,12 @@ import { log } from "../utils/cli.ts";
 import { addInstructions } from "./instructions.ts";
 import { agent, installFromNpmTarball, setupProcessAgentEnv } from "./shared.ts";
 
-// model configuration based on effort level
-// uses model family aliases that auto-resolve to latest version
-const codexModels: Record<Effort, string> = {
-  nothink: "gpt-4o-mini",
-  think: "gpt-5.2-instant",
-  max: "gpt-5.2-thinking",
+// reasoning effort configuration based on effort level
+// uses model_reasoning_effort parameter for o1 models
+const codexReasoningEffort: Record<Effort, string | undefined> = {
+  nothink: "low",
+  think: undefined, // use default
+  max: "xhigh",
 } as const;
 
 interface WriteCodexConfigParams {
@@ -88,9 +88,13 @@ export const codex = agent({
       CODEX_HOME: codexDir, // point Codex to our config directory
     });
 
-    // get model based on effort level
-    const model = codexModels[effort];
-    log.info(`Using model: ${model}`);
+    // get reasoning effort based on effort level
+    const modelReasoningEffort = codexReasoningEffort[effort];
+    if (modelReasoningEffort) {
+      log.info(`Using model_reasoning_effort: ${modelReasoningEffort}`);
+    } else {
+      log.info(`Using default reasoning effort`);
+    }
 
     // Configure Codex
     const codexOptions: CodexOptions = {
@@ -103,22 +107,26 @@ export const codex = agent({
     }
 
     const codex = new Codex(codexOptions);
-    const thread = codex.startThread(
-      payload.sandbox
-        ? {
-            model,
-            approvalPolicy: "never",
-            sandboxMode: "read-only",
-            networkAccessEnabled: false,
-          }
-        : {
-            model,
-            approvalPolicy: "never",
-            // use danger-full-access to allow git operations (workspace-write blocks .git directory writes)
-            sandboxMode: "danger-full-access",
-            networkAccessEnabled: true,
-          }
-    );
+
+    // Build thread options with optional model_reasoning_effort
+    const baseThreadOptions = payload.sandbox
+      ? {
+          approvalPolicy: "never" as const,
+          sandboxMode: "read-only" as const,
+          networkAccessEnabled: false,
+        }
+      : {
+          approvalPolicy: "never" as const,
+          // use danger-full-access to allow git operations (workspace-write blocks .git directory writes)
+          sandboxMode: "danger-full-access" as const,
+          networkAccessEnabled: true,
+        };
+
+    const threadOptions = modelReasoningEffort
+      ? { ...baseThreadOptions, model_reasoning_effort: modelReasoningEffort }
+      : baseThreadOptions;
+
+    const thread = codex.startThread(threadOptions);
 
     try {
       const streamedTurn = await thread.runStreamed(addInstructions({ payload, repo }));

agents/gemini.ts

@@ -12,11 +12,11 @@ import {
 } from "./shared.ts";
 
 // model configuration based on effort level
-// uses model family aliases that auto-resolve to latest version
+// auto for default, flash for fast/low-cost, pro for max capability
 const geminiModels = {
-  nothink: "gemini-2.5-pro",
-  think: "gemini-3-pro",
-  max: "gemini-3-pro",
+  nothink: "gemini-2.5-flash",
+  think: "gemini-2.5-pro", // auto selection
+  max: "gemini-2.5-pro",
 } as const;
 
 // gemini cli event types inferred from stream-json output (NDJSON format)

agents/opencode.ts

@@ -68,7 +68,7 @@ export const opencode = agent({
 
     // add API keys from apiKeys object
     for (const [key, value] of Object.entries(apiKeys || {})) {
-      env[key] = value;
+      env[key.toUpperCase()] = value;
       // also set GOOGLE_GENERATIVE_AI_API_KEY for Google provider compatibility
       if (key === "GEMINI_API_KEY") {
         env.GOOGLE_GENERATIVE_AI_API_KEY = value;

main.ts

@@ -401,10 +401,10 @@ function parsePayload(inputs: Inputs): Payload {
     if (!("~pullfrog" in parsedPrompt)) {
       throw new Error();
     }
-    // internal invocation: use effort from payload, fallback to input
+    // internal invocation: use effort from payload, fallback to input, default to "think"
     return {
       ...parsedPrompt,
-      effort: parsedPrompt.effort ?? inputs.effort,
+      effort: parsedPrompt.effort ?? inputs.effort ?? "think",
     } as Payload;
   } catch {
     // external invocation: use effort from input