Merge pull request #1243 from Steve0x2a/fix/omit-temperature-gpt-6-responses

fix: omit temperature for gpt-6 under openai-responses

electron/services/agentMemorySynthesisService.ts

@@ -10,6 +10,7 @@ import {
 } from './agentMemoryService'
 import { createAgentTextRedactionSession } from './agentTextRedaction'
 import { filterAutomaticMemoryEntries, filterAutomaticMemoryPatch, type AgentMemoryQualityIssue } from './agentMemoryQuality'
+import { omitsTemperature } from './agentSamplingOptions'
 import type { AgentFeedbackMemoryEvidence, AgentFeedbackMemorySample, AgentFeedbackReason } from './agentFeedbackStore'
 import type { AgentModelConfig } from './agentService'
 
@@ -66,9 +67,7 @@ function latestUserText(messages: unknown[]): string {
 }
 
 function memorySamplingOptions(config: AgentModelConfig): { temperature?: number } {
-  const modelId = String(config.model || '').trim().toLocaleLowerCase()
-  const reasoningModel = config.protocol === 'openai-responses' && /^(?:gpt-5|o[1-4](?:-|$))/.test(modelId)
-  return reasoningModel || config.protocol === 'anthropic' ? {} : { temperature: 0.1 }
+  return omitsTemperature(config) ? {} : { temperature: 0.1 }
 }
 
 function modelIsConfigured(config?: AgentModelConfig | null): config is AgentModelConfig {

electron/services/agentModelLedService.ts

@@ -59,6 +59,7 @@ import {
 } from './agentMemoryRuntimeGuidance'
 import { createAgentModel } from './agentModelAdapter'
 import { truncateStringToSchemaMax } from './agentModelSafety'
+import { omitsTemperature } from './agentSamplingOptions'
 import {
   AgentRunStore,
   assessAgentRunResume,
@@ -851,9 +852,7 @@ function providerOptions(config: AgentModelConfig, promptCacheKey: string): Prov
 }
 
 function samplingOptions(config: AgentModelConfig, temperature: number): { temperature?: number } {
-  const model = text(config.model).toLocaleLowerCase()
-  if (config.protocol === 'anthropic' || (config.protocol === 'openai-responses' && /^(?:gpt-5|o[1-4](?:-|$))/.test(model))) return {}
-  return { temperature }
+  return omitsTemperature(config) ? {} : { temperature }
 }
 
 function compactToolOutputForStorage(toolName: string, output: unknown): unknown {

electron/services/agentSamplingOptions.ts

@@ -0,0 +1,12 @@
+export type AgentSamplingConfig = {
+  protocol?: string
+  model?: string
+}
+
+const OPENAI_RESPONSES_REASONING_MODEL = /^(?:gpt-(?:5|6)|o[1-4](?:-|$))/
+
+export function omitsTemperature(config: AgentSamplingConfig): boolean {
+  if (config.protocol === 'anthropic') return true
+  const modelId = String(config.model || '').trim().toLocaleLowerCase()
+  return config.protocol === 'openai-responses' && OPENAI_RESPONSES_REASONING_MODEL.test(modelId)
+}

electron/services/agentTitleService.ts

@@ -1,5 +1,6 @@
 import { streamText } from 'ai'
 import { createAgentModel } from './agentModelAdapter'
+import { omitsTemperature } from './agentSamplingOptions'
 import { createAgentTextRedactionSession } from './agentTextRedaction'
 import type { AgentModelConfig } from './agentService'
 
@@ -19,11 +20,8 @@ function fallbackTitle(value: string): string {
 
 function titleSamplingOptions(config: AgentModelConfig): { temperature?: number } {
   const modelId = String(config.model || '').trim().toLocaleLowerCase()
-  const openAiReasoningModel = config.protocol === 'openai-responses'
-    && /^(?:gpt-5|o[1-4](?:-|$))/.test(modelId)
-  const nativeAnthropicModel = config.protocol === 'anthropic'
   const modernClaudeModel = /(?:^|\/)claude-(?:opus-4-(?:[7-9]|\d{2,})|(?:sonnet|opus|haiku|fable|mythos)-(?:[5-9]|\d{2,})(?:-|$))/.test(modelId)
-  return openAiReasoningModel || nativeAnthropicModel || modernClaudeModel ? {} : { temperature: 0.2 }
+  return omitsTemperature(config) || modernClaudeModel ? {} : { temperature: 0.2 }
 }
 
 function titleErrorMessage(error: unknown): string {

scripts/check-agent-sampling-options.cjs

@@ -0,0 +1,138 @@
+// Exercise the production omit-temperature predicate without loading Electron or native modules.
+const assert = require('node:assert/strict')
+const fs = require('node:fs')
+const path = require('node:path')
+const vm = require('node:vm')
+const { test } = require('node:test')
+
+const root = path.join(__dirname, '..')
+const helperPath = path.join(root, 'electron/services/agentSamplingOptions.ts')
+const ledPath = path.join(root, 'electron/services/agentModelLedService.ts')
+const titlePath = path.join(root, 'electron/services/agentTitleService.ts')
+const memoryPath = path.join(root, 'electron/services/agentMemorySynthesisService.ts')
+
+function extractFunction(source, name) {
+  const header = new RegExp(`function ${name}\\([^)]*\\)(?:\\s*:\\s*\\{[^}]*\\})?\\s*\\{`)
+  const match = header.exec(source)
+  assert.ok(match, `missing function ${name}`)
+  const start = match.index
+  const brace = start + match[0].length - 1
+  let depth = 0
+  for (let index = brace; index < source.length; index += 1) {
+    const char = source[index]
+    if (char === '{') depth += 1
+    else if (char === '}') {
+      depth -= 1
+      if (depth === 0) return source.slice(start, index + 1)
+    }
+  }
+  throw new Error(`unterminated function ${name}`)
+}
+
+function stripTypes(source) {
+  return source
+    .replace(/\bexport type [A-Za-z0-9_]+ = \{[\s\S]*?\n\}\n/, '')
+    .replace(/\bexport /g, '')
+    .replace(/: Agent(?:Sampling|Model)Config/g, '')
+    .replace(/: \{ protocol\?: string; model\?: string \}/g, '')
+    .replace(/temperature: number/g, 'temperature')
+    .replace(/\): boolean/g, ')')
+    .replace(/\): \{ temperature\?: number \}/g, ')')
+}
+
+function loadHelper() {
+  const source = fs.readFileSync(helperPath, 'utf8')
+  const js = stripTypes(source)
+  const module = { exports: {} }
+  vm.runInNewContext(`${js}\nmodule.exports = { omitsTemperature }`, {
+    module,
+    exports: module.exports,
+  })
+  return module.exports
+}
+
+function loadOmitsTemperature() {
+  if (fs.existsSync(helperPath)) return loadHelper().omitsTemperature
+
+  const led = extractFunction(fs.readFileSync(ledPath, 'utf8'), 'samplingOptions')
+  const title = extractFunction(fs.readFileSync(titlePath, 'utf8'), 'titleSamplingOptions')
+  const memory = extractFunction(fs.readFileSync(memoryPath, 'utf8'), 'memorySamplingOptions')
+  const js = `
+    function text(value) { return String(value ?? '').replace(/\\u0000/g, '').trim() }
+    ${stripTypes(led)}
+    ${stripTypes(title)}
+    ${stripTypes(memory)}
+    function omitsTemperature(config) {
+      return !('temperature' in samplingOptions(config, 0.7))
+        && !('temperature' in titleSamplingOptions(config))
+        && !('temperature' in memorySamplingOptions(config))
+    }
+  `
+  const module = { exports: {} }
+  vm.runInNewContext(`${js}\nmodule.exports = { omitsTemperature }`, { module, exports: module.exports })
+  return module.exports.omitsTemperature
+}
+
+function loadSamplingOptions() {
+  const omits = loadOmitsTemperature()
+  return (config, temperature) => (omits(config) ? {} : { temperature })
+}
+
+const omitsTemperature = loadOmitsTemperature()
+const samplingOptions = loadSamplingOptions()
+const plain = (value) => JSON.parse(JSON.stringify(value))
+
+function responses(model) {
+  return { protocol: 'openai-responses', model }
+}
+
+test('openai-responses gpt-6-luna omits temperature', () => {
+  assert.equal(omitsTemperature(responses('gpt-6-luna')), true)
+  assert.deepEqual(plain(samplingOptions(responses('gpt-6-luna'), 0.7)), {})
+})
+
+test('openai-responses gpt-6-* variants omit temperature', () => {
+  for (const model of ['gpt-6', 'gpt-6-something', 'GPT-6-Luna']) {
+    assert.equal(omitsTemperature(responses(model)), true, model)
+    assert.deepEqual(plain(samplingOptions(responses(model), 0.2)), {})
+  }
+})
+
+test('openai-responses gpt-5 and o-series still omit temperature', () => {
+  for (const model of ['gpt-5', 'gpt-5.5', 'gpt-5.4-mini', 'o1', 'o1-preview', 'o3', 'o4-mini']) {
+    assert.equal(omitsTemperature(responses(model)), true, model)
+    assert.deepEqual(plain(samplingOptions(responses(model), 0.7)), {})
+  }
+})
+
+test('non-reasoning openai-responses still includes temperature', () => {
+  for (const model of ['gpt-4o', 'gpt-4.1', 'chatgpt-4o-latest']) {
+    assert.equal(omitsTemperature(responses(model)), false, model)
+    assert.deepEqual(plain(samplingOptions(responses(model), 0.7)), { temperature: 0.7 })
+  }
+})
+
+test('openai-compatible gpt-6-luna still includes temperature', () => {
+  const config = { protocol: 'openai-compatible', model: 'gpt-6-luna' }
+  assert.equal(omitsTemperature(config), false)
+  assert.deepEqual(plain(samplingOptions(config, 0.3)), { temperature: 0.3 })
+})
+
+test('anthropic path still omits temperature for any model', () => {
+  for (const model of ['claude-sonnet-4', 'gpt-4o', 'gpt-6-luna']) {
+    const config = { protocol: 'anthropic', model }
+    assert.equal(omitsTemperature(config), true, model)
+    assert.deepEqual(plain(samplingOptions(config, 0.7)), {})
+  }
+})
+
+test('agent chat, title, and memory sampling share omitsTemperature', () => {
+  if (!fs.existsSync(helperPath)) {
+    assert.fail('shared helper electron/services/agentSamplingOptions.ts is required')
+  }
+  for (const file of [ledPath, titlePath, memoryPath]) {
+    const source = fs.readFileSync(file, 'utf8')
+    assert.match(source, /from '\.\/agentSamplingOptions'/)
+    assert.match(source, /omitsTemperature\s*\(/)
+  }
+})