Merge branch 'main' into eval-claude-codex

Author: jahooma

.agents/base2/base2-editor.ts

@@ -0,0 +1,8 @@
+import { createBase2 } from './base2'
+
+const definition = {
+  ...createBase2('default', { useEditor: true }),
+  id: 'base2-editor',
+  displayName: 'Buffy the Editor Orchestrator',
+}
+export default definition

.agents/base2/base2.ts

@@ -11,9 +11,14 @@ export function createBase2(
   options?: {
     hasNoValidation?: boolean
     planOnly?: boolean
+    useEditor?: boolean
   },
 ): Omit<SecretAgentDefinition, 'id'> {
-  const { hasNoValidation = mode === 'fast', planOnly = false } = options ?? {}
+  const {
+    hasNoValidation = mode === 'fast',
+    planOnly = false,
+    useEditor = false,
+  } = options ?? {}
   const isDefault = mode === 'default'
   const isFast = mode === 'fast'
   const isMax = mode === 'max'
@@ -65,6 +70,7 @@ export function createBase2(
       'researcher-docs',
       isLite ? 'commander-lite' : 'commander',
       isLite && 'editor-gpt-5',
+      useEditor && 'editor',
       isMax && 'editor-best-of-n-max',
       isMax && 'thinker-best-of-n-opus',
       !isLite && 'code-reviewer-opus',
@@ -119,6 +125,8 @@ Use the spawn_agents tool to spawn specialized agents to help you complete the u
     '- Spawn context-gathering agents (file pickers, code-searcher, directory-lister, glob-matcher, and web/docs researchers) before making edits.',
     isLite &&
       '- Spawn the editor-gpt-5 agent to implement the changes after you have gathered all the context you need.',
+    useEditor &&
+      '- Spawn the editor agent to implement the changes after you have gathered all the context you need.',
     isMax &&
       '- Spawn the thinker-best-of-n-opus after gathering context to solve complex problems.',
     isMax &&
@@ -171,11 +179,13 @@ ${buildArray(
 [ You read a few other relevant files using the read_files tool ]
 
 ${
-  isDefault || isFast
-    ? '[ You implement the changes using the str_replace or write_file tools ]'
-    : isLite
-      ? '[ You implement the changes using the editor-gpt-5 agent ]'
-      : '[ You implement the changes using the editor-best-of-n-max agent ]'
+  useEditor
+    ? `[ You implement the changes using the editor agent ]`
+    : isDefault || isFast
+      ? '[ You implement the changes using the str_replace or write_file tools ]'
+      : isLite
+        ? '[ You implement the changes using the editor-gpt-5 agent ]'
+        : '[ You implement the changes using the editor-best-of-n-max agent ]'
 }
 
 ${
@@ -225,6 +235,7 @@ ${PLACEHOLDER.GIT_CHANGES_PROMPT}
           isMax,
           isLite,
           hasNoValidation,
+          useEditor,
         }),
     stepPrompt: planOnly
       ? buildPlanOnlyStepPrompt({})
@@ -233,6 +244,7 @@ ${PLACEHOLDER.GIT_CHANGES_PROMPT}
           isMax,
           hasNoValidation,
           isSonnet,
+          useEditor,
         }),
 
     handleSteps: function* ({ params }) {
@@ -265,13 +277,15 @@ function buildImplementationInstructionsPrompt({
   isMax,
   isLite,
   hasNoValidation,
+  useEditor,
 }: {
   isSonnet: boolean
   isFast: boolean
   isDefault: boolean
   isMax: boolean
   isLite: boolean
   hasNoValidation: boolean
+  useEditor: boolean
 }) {
   return `Act as a helpful assistant and freely respond to the user's request however would be most helpful to the user. Use your judgement to orchestrate the completion of the user's request using your specialized sub-agents and tools as needed. Take your time and be comprehensive. Don't surprise the user. For example, don't modify files if the user has not asked you to do so at least implicitly.
 
@@ -287,8 +301,10 @@ ${buildArray(
     `- For any task requiring 3+ steps, use the write_todos tool to write out your step-by-step implementation plan. Include ALL of the applicable tasks in the list.${isFast ? '' : ' You should include a step to review the changes after you have implemented the changes.'}:${hasNoValidation ? '' : ' You should include at least one step to validate/test your changes: be specific about whether to typecheck, run tests, run lints, etc.'} You may be able to do reviewing and validation in parallel in the same step. Skip write_todos for simple tasks like quick edits or answering questions.`,
   isLite &&
     '- IMPORTANT: You must spawn the editor-gpt-5 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 changes.',
+  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 &&
@@ -308,15 +324,19 @@ function buildImplementationStepPrompt({
   isMax,
   hasNoValidation,
   isSonnet,
+  useEditor,
 }: {
   isFast: boolean
   isMax: boolean
   hasNoValidation: boolean
   isSonnet: boolean
+  useEditor: boolean
 }) {
   return buildArray(
     isMax &&
       `Keep working until the user's request is completely satisfied${!hasNoValidation ? ' and validated' : ''}, or until you require more information from the user.`,
+    useEditor &&
+      `You must spawn the 'editor' agent to implement code changes, since it will do the best job of implementing the changes.`,
     isMax &&
       `You must spawn the 'editor-best-of-n-max' agent to implement code changes, since it will generate the best code changes.`,
     isMax && 'Spawn the thinker-best-of-n-opus to solve complex problems.',

.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,
@@ -245,12 +245,14 @@ function* handleStepsMax({
   // Model selection pattern for max mode, using opus and gpt-5
   const MAX_MODEL_PATTERN = [
     'editor-implementor-opus',
-    'editor-implementor-gemini',
+    'editor-implementor-opus',
+    // 'editor-implementor-gemini',
     'editor-implementor-gpt-5',
     'editor-implementor-opus',
     'editor-implementor-opus',
     'editor-implementor-gpt-5',
-    'editor-implementor-gemini',
+    // 'editor-implementor-gemini',
+    'editor-implementor-opus',
     'editor-implementor-opus',
     'editor-implementor-opus',
     'editor-implementor-opus',
@@ -296,6 +298,10 @@ 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 +310,9 @@ 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 +433,14 @@ 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 +460,6 @@ 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