Update best of n selector, editor implementor to think before selecting/implementing

Author: jahooma

.agents/base2/base2.ts

@@ -304,7 +304,7 @@ ${buildArray(
   useEditor &&
     '- IMPORTANT: You must spawn the editor agent to implement the changes after you have gathered all the context you need. This agent will do the best job of implementing the changes so you must spawn it for all non-trivial changes. Do not pass any prompt or params to the editor agent when spawning it. It will make its own best choices of what to do.',
   isMax &&
-    `- IMPORTANT: You must spawn the editor-best-of-n-max agent to implement non-trivial code changes, since it will generate the best code changes from multiple implementation proposals. This is the best way to make high quality code changes -- strongly prefer using this agent over the str_replace or write_file tools, unless the change is very straightforward and obvious.`,
+    `- IMPORTANT: You must spawn the editor-best-of-n-max agent to implement non-trivial code changes, since it will generate the best code changes from multiple implementation proposals. This is the best way to make high quality code changes -- strongly prefer using this agent over the str_replace or write_file tools, unless the change is very straightforward and obvious. Do not pass any prompt or params to the editor agent when spawning it. It will make its own best choices of what to do.`,
   (isDefault || isFast) &&
     '- Implement the changes using the str_replace or write_file tools.',
   isFast &&

.agents/editor/best-of-n/best-of-n-selector.ts

@@ -26,11 +26,6 @@ export const createBestOfNSelector = (options: {
         effort: 'high',
       },
     }),
-    ...(isOpus && {
-      reasoningOptions: {
-        max_tokens: 4000,
-      },
-    }),
     displayName: isGpt5
       ? 'Best-of-N GPT-5 Implementation Selector'
       : isGemini
@@ -114,12 +109,10 @@ Try to select an implementation that fulfills all the requirements in the user's
 ## Response Format
 
 ${
-  isSonnet
-    ? `Use <think> tags to briefly consider the implementations as needed to pick the best implementation.
-
-If the best one is obvious or the implementations are very similar, you may not need to think very much (a few words suffice) or you may not need to use think tags at all, just pick the best one and output it. You have a dual goal of picking the best implementation and being fast (using as few words as possible).
+  isSonnet || isOpus
+    ? `Use <think> tags to write out your thoughts about the implementations as needed to pick the best implementation. IMPORTANT: You should think really really hard to make sure you pick the absolute best implementation! As soon as you know for sure which implementation is the best, you should output your choice.
 
-Then, do not write any other explanations AT ALL. You should directly output a single tool call to set_output with the selected implementationId.`
+Then, do not write any other explanations AT ALL. You should directly output a single tool call to set_output with the selected implementationId and short reason.`
     : `Output a single tool call to set_output with the selected implementationId. Do not write anything else.`
 }`,
   }

.agents/editor/best-of-n/editor-best-of-n.ts

@@ -52,7 +52,7 @@ export function createBestOfNEditor(
         properties: {
           n: {
             type: 'number',
-            description: `Number of parallel implementor agents to spawn. Defaults to ${isDefault ? 4 : 5}. Use fewer for simple tasks and max of 10 for complex tasks.`,
+            description: `Number of parallel implementor agents to spawn. Defaults to ${isMax ? 4 : 3}. Use fewer for simple tasks and max of 10 for complex tasks.`,
           },
         },
       },
@@ -73,7 +73,7 @@ function* handleStepsDefault({
 }: AgentStepContext): ReturnType<
   NonNullable<SecretAgentDefinition['handleSteps']>
 > {
-  const DEFAULT_N = 4
+  const DEFAULT_N = 3
   const selectorAgent = 'best-of-n-selector'
   const n = Math.min(
     10,
@@ -235,7 +235,7 @@ function* handleStepsMax({
 }: AgentStepContext): ReturnType<
   NonNullable<SecretAgentDefinition['handleSteps']>
 > {
-  const MAX_N = 5
+  const MAX_N = 4
   const selectorAgent = 'best-of-n-selector-opus'
   const n = Math.min(
     10,
@@ -296,6 +296,10 @@ function* handleStepsMax({
     implementorResults,
   ) as any[]
 
+  logger.info(
+    { implementorResults, spawnedImplementations },
+    'spawnedImplementations',
+  )
   // Extract all the plans from the structured outputs
   const letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZ'
   // Parse implementations from spawn results
@@ -304,14 +308,9 @@ function* handleStepsMax({
     content:
       'errorMessage' in result
         ? `Error: ${result.errorMessage}`
-        : extractLastMessageText(result),
+        : extractLastMessageText(result) ?? '',
   }))
 
-  logger.info(
-    { spawnedImplementations, implementations },
-    'spawnedImplementations',
-  )
-
   // Spawn selector with implementations as params
   const { toolResult: selectorResult, agentState: selectorAgentState } = yield {
     toolName: 'spawn_agents',
@@ -432,14 +431,14 @@ function* handleStepsOpus({
 }: AgentStepContext): ReturnType<
   NonNullable<SecretAgentDefinition['handleSteps']>
 > {
-  const DEFAULT_N = 5
+  const DEFAULT_N = 3
   const selectorAgent = 'best-of-n-selector-opus'
   const n = Math.min(
     10,
     Math.max(1, (params?.n as number | undefined) ?? DEFAULT_N),
   )
 
-  // Spawn implementor agents: 1 gemini + rest sonnet (if n >= 2)
+  // Spawn implementor agents
   const implementorAgents = []
   for (let i = 0; i < n; i++) {
     implementorAgents.push({
@@ -459,8 +458,6 @@ function* handleStepsOpus({
   const spawnedImplementations =
     extractSpawnResults<{ text: string }[]>(implementorResults)
 
-  logger.info({ spawnedImplementations }, 'spawnedImplementations')
-
   // Extract all the plans from the structured outputs
   const letters = 'ABCDEFGHIJKLMNOPQRSTUVWXYZ'
   // Parse implementations from spawn results

.agents/editor/best-of-n/editor-implementor.ts

@@ -72,12 +72,14 @@ ${
   isGpt5 || isGemini
     ? ``
     : `
-You can also use <think> tags interspersed between tool calls to think about the best way to implement the changes. Keep these thoughts very brief. You may not need to use think tags at all.
+IMPORTANT: Before you start writing your implementation, you should use <think> tags to think about the best way to implement the changes. You should think really really hard to make sure you implement the changes in the best way possible. Take as much time as you to think through all the cases to produce the best changes.
+
+You can also use <think> tags interspersed between tool calls to think about the best way to implement the changes.
 
 <example>
 
 <think>
-[ Thoughts about the best way to implement the feature ]
+[ Long think about the best way to implement the changes ]
 </think>
 
 <codebuff_tool_call>
@@ -99,7 +101,7 @@ You can also use <think> tags interspersed between tool calls to think about the
 </example>`
 }
 
-After the edit tool calls, you can optionally mention any follow-up steps to take, like deleting a file, or a sepcific way to validate the changes. There's no need to use the set_output tool as your entire response will be included in the output.
+After the edit tool calls, you can optionally mention any follow-up steps to take, like deleting a file, or a specific way to validate the changes. There's no need to use the set_output tool as your entire response will be included in the output.
 
 Your implementation should:
 - Be complete and comprehensive