@shawnstack/quickforge 1.7.4 → 1.7.5

package/server/agent-manager.mjs

@@ -13,6 +13,7 @@ import {
   formatSubagentTask,
 } from './subagents.mjs'
 import { agentProfileSnapshot, getAgentProfile } from './agent-profiles.mjs'
+import { modelBindingFromModel, resolveImplicitModelPreference, resolveModelBinding } from './model-catalog.mjs'
 import { agentProfileFromMarkdown } from './agent-profile-files.mjs'
 import {
   applyCapabilityPolicy,
@@ -200,7 +201,7 @@ function hasFullAccess(session) {
 
 const IDLE_TIMEOUT_MS = 30 * 60 * 1000 // 30 minutes
 const ABORT_IDLE_WAIT_TIMEOUT_MS = 3000
-const SUBAGENT_DEFAULT_TIMEOUT_MS = 30 * 60 * 1000
+const SUBAGENT_DEFAULT_TIMEOUT_MS = 60 * 60 * 1000
 const SUBAGENT_TRACE_THROTTLE_MS = 150
 
 /**
@@ -515,6 +516,37 @@ function updateSessionMessages(session, messages) {
   session.agent.state.messages = messages
 }
 
+const CLIENT_MESSAGE_ID_FIELD = 'quickforgeClientMessageId'
+
+function isManagedCloudModel(model) {
+  return model?.provider === 'quickforge-cloud' && model?.quickforgeModelSource === 'cloud'
+}
+
+function objectMetadata(message) {
+  const metadata = message?.metadata
+  return metadata && typeof metadata === 'object' && !Array.isArray(metadata) ? metadata : {}
+}
+
+function logicalMessageId(message) {
+  const value = objectMetadata(message)[CLIENT_MESSAGE_ID_FIELD]
+  return typeof value === 'string' && /^[A-Za-z0-9._:-]{1,128}$/.test(value) ? value : undefined
+}
+
+function messageWithLogicalId(message) {
+  if (!message || typeof message !== 'object' || logicalMessageId(message)) return message
+  return {
+    ...message,
+    metadata: {
+      ...objectMetadata(message),
+      [CLIENT_MESSAGE_ID_FIELD]: `qfcm_${randomUUID()}`,
+    },
+  }
+}
+
+function prepareCloudUserMessage(session, message) {
+  return isManagedCloudModel(session?.model) ? messageWithLogicalId(message) : message
+}
+
 function resetSessionCompaction(session) {
   session.contextCompaction = null
   session.lastAutoCompactAt = null
@@ -627,6 +659,9 @@ async function summarySession(session, initialUserMessage, summaryOptions) {
       accessMode: session.accessMode,
       yoloMode: session.yoloMode,
       model: session.model,
+      modelRef: session.modelRef,
+      modelAccessContext: session.modelAccessContext,
+      resolvePersistedModel: true,
       thinkingLevel: session.thinkingLevel,
       messages: compactedMessages,
       title: compactedTitle,
@@ -1081,7 +1116,7 @@ model:
   modelId: ${frontmatterString(spec.model.modelId)}${spec.model.api ? `
   api: ${frontmatterString(spec.model.api)}` : ''}${spec.model.baseUrl ? `
   baseUrl: ${frontmatterString(spec.model.baseUrl)}` : ''}` : ''}
-max-runtime-ms: ${Math.max(1000, Math.min(Number(spec.maxRuntimeMs || SUBAGENT_DEFAULT_TIMEOUT_MS), 30 * 60 * 1000))}
+max-runtime-ms: ${Math.max(1000, Math.min(Number(spec.maxRuntimeMs || SUBAGENT_DEFAULT_TIMEOUT_MS), 60 * 60 * 1000))}
 max-tool-calls: ${Math.max(1, Math.min(Number(spec.maxToolCalls || 300), 300))}
 ---
 ${spec.systemPrompt}
@@ -1135,10 +1170,15 @@ async function runSubagent(parentSession, params, parentSignal, onUpdate) {
     throw new Error('No active model is configured for the parent session.')
   }
 
-  const timeoutMs = Math.max(1000, Math.min(Number(definition.maxRuntimeMs || SUBAGENT_DEFAULT_TIMEOUT_MS), 30 * 60 * 1000))
+  const timeoutMs = Math.max(1000, Math.min(Number(definition.maxRuntimeMs || SUBAGENT_DEFAULT_TIMEOUT_MS), 60 * 60 * 1000))
   definition.capabilityPolicy ||= inferCapabilityPolicy(definition.allowedTools || [])
   definition.allowedTools = applyCapabilityPolicy(definition.allowedTools || [], definition.capabilityPolicy)
-  const { model: subagentModel, info: subagentModelInfo } = await resolveAgentProfileModel(definition, parentSession.model, readStore)
+  const { model: subagentModel, info: subagentModelInfo } = await resolveAgentProfileModel(
+    definition,
+    parentSession.model,
+    readStore,
+    parentSession.modelAccessContext || {},
+  )
   const subagentThinkingLevel = resolveAgentProfileThinkingLevel(definition, parentSession.thinkingLevel, subagentModel)
   const subagentSessionId = `${parentSession.sessionId}:subagent:${definition.name}:${randomUUID()}`
   const startedAt = Date.now()
@@ -1511,6 +1551,9 @@ export async function createAgent(sessionId, config = {}) {
     accessMode: rawAccessMode,
     yoloMode = false,
     model = null,
+    modelRef = null,
+    modelAccessContext = null,
+    resolvePersistedModel = false,
     thinkingLevel = 'off',
     messages = [],
     systemPrompt = null,
@@ -1555,24 +1598,36 @@ export async function createAgent(sessionId, config = {}) {
 
   // Resolve model
   let resolvedModel = model
+  let resolvedModelRef = modelRef
   if (!resolvedModel) {
-    // Try to load from storage
+    // Try to load the active preference from storage. A stale/hidden implicit
+    // preference may fall back, while an explicit session binding never does.
     try {
       const settings = await readStore('settings')
       const raw = settings?.['active-model']
-      if (raw) resolvedModel = typeof raw === 'string' ? JSON.parse(raw) : raw
+      resolvedModel = await resolveImplicitModelPreference(raw, modelAccessContext || {})
     } catch {
       // ignore
     }
   }
   if (agentProfile?.model?.mode === 'fixed') {
-    resolvedModel = (await resolveAgentProfileModel(agentProfile, resolvedModel, readStore)).model
+    const profileBinding = await resolveAgentProfileModel(agentProfile, resolvedModel, readStore, modelAccessContext || {})
+    resolvedModel = profileBinding.model
+    resolvedModelRef = profileBinding.modelRef || null
     resolvedAgentProfile = {
       ...agentProfile,
       model: modelReferenceSnapshot(agentProfile.model),
     }
   }
   if (resolvedAgentProfile && !resolvedModel) throw new Error('No active model is configured for the agent session.')
+  const resolvedBinding = resolvedModel
+    ? (resolvedModelRef
+        ? { model: resolvedModel, modelRef: resolvedModelRef }
+        : model
+          ? { model: resolvedModel, modelRef: null }
+          : await modelBindingFromModel(resolvedModel))
+    : { model: resolvedModel, modelRef: null }
+  resolvedModel = resolvedBinding.model
   const resolvedThinkingLevel = agentProfile
     ? resolveAgentProfileThinkingLevel(agentProfile, thinkingLevel, resolvedModel)
     : thinkingLevel
@@ -1682,6 +1737,9 @@ export async function createAgent(sessionId, config = {}) {
     accessMode,
     yoloMode: resolvedYoloMode,
     model: resolvedModel,
+    modelRef: resolvedBinding.modelRef,
+    modelAccessContext: modelAccessContext || null,
+    resolvePersistedModel: resolvePersistedModel === true || Boolean(resolvedBinding.modelRef),
     thinkingLevel: resolvedThinkingLevel,
     scope,
     title,
@@ -1789,14 +1847,15 @@ export async function createAgent(sessionId, config = {}) {
     }
 
     if (event.type === 'message_end') {
-      const isInitialUserMessage = (event.message?.role === 'user' || event.message?.role === 'user-with-attachments')
-        && session.agent.state.messages.length === 1
-      if (isInitialUserMessage) {
-        // Persist the first user message before announcing the session or
-        // starting the asynchronous AI title request.
+      const isUserMessage = event.message?.role === 'user' || event.message?.role === 'user-with-attachments'
+      const isInitialUserMessage = isUserMessage && session.agent.state.messages.length === 1
+      const requiresDurableCloudMessage = isUserMessage && isManagedCloudModel(session.model) && Boolean(logicalMessageId(event.message))
+      if (isInitialUserMessage || requiresDurableCloudMessage) {
+        // Persist every managed Cloud user message before the provider stream starts,
+        // so its logical ID survives a process restart and resolves to the same private key.
         try {
           const metadata = await flushSessionPersist(session)
-          if (metadata) {
+          if (metadata && isInitialUserMessage) {
             if (!session.sessionCreatedEmitted) {
               session.sessionCreatedEmitted = true
               emitSessionEvent(session, { type: 'session_created', metadata })
@@ -1804,7 +1863,7 @@ export async function createAgent(sessionId, config = {}) {
             scheduleSessionTitleGeneration(session, event.message)
           }
         } catch (err) {
-          logger.error(`Failed to persist initial user message for session ${sessionId}:`, err, { sessionId })
+          logger.error(`Failed to persist user message for session ${sessionId}:`, err, { sessionId })
         }
       } else {
         // Debounced persist for crash recovery; coalesces the many message_end
@@ -1848,7 +1907,7 @@ function sessionLastModifiedFromMessages(messages, fallback) {
  * Persist session data to storage.
  */
 async function persistSessionUnlocked(session) {
-  const { sessionId, agent, scope, projectId, source, channelId, channelName, title, titleSource, createdAt, lastModified: storedLastModified, status, startedAt, finishedAt, model, thinkingLevel, accessMode, yoloMode, contextCompaction } = session
+  const { sessionId, agent, scope, projectId, source, channelId, channelName, title, titleSource, createdAt, lastModified: storedLastModified, status, startedAt, finishedAt, model, modelRef, thinkingLevel, accessMode, yoloMode, contextCompaction } = session
   const messages = agent.state.messages
 
   if (messages.length === 0) {
@@ -1871,6 +1930,7 @@ async function persistSessionUnlocked(session) {
     title,
     titleSource,
     model,
+    modelRef: modelRef || undefined,
     thinkingLevel,
     accessMode,
     yoloMode,
@@ -2063,7 +2123,10 @@ export async function rollbackSessionMessages(sessionId, rollbackMessageIndex) {
     throw Object.assign(new Error('Session not found'), { statusCode: 404 })
   }
   if (session.agent.state.isStreaming) {
-    throw Object.assign(new Error('Generation is still running. Stop it or wait until it finishes before rolling back.'), { statusCode: 409 })
+    throw Object.assign(new Error('Generation is still running. Stop it or wait until it finishes before rolling back.'), {
+      statusCode: 409,
+      errorCode: 'GENERATION_STILL_RUNNING_BEFORE_ROLLBACK',
+    })
   }
 
   const messages = Array.isArray(session.agent.state.messages) ? session.agent.state.messages : []
@@ -2127,18 +2190,42 @@ function scheduleSessionTitleGeneration(session, userMessage) {
   session.titleGenerationPromiseId = generationId
 }
 
+async function refreshSessionModelBinding(session) {
+  if (!session?.model || (!session.modelRef && session.resolvePersistedModel !== true)) return session
+  const binding = await resolveModelBinding(
+    session.modelRef ? { modelRef: session.modelRef } : { model: session.model },
+    {
+      context: session.modelAccessContext || {},
+      currentModel: session.model,
+      allowCurrentHidden: true,
+      forExecution: true,
+      legacySnapshot: session.model,
+    },
+  )
+  session.model = binding.model
+  session.modelRef = binding.modelRef
+  session.resolvePersistedModel = true
+  session.agent.state.model = binding.model
+  return session
+}
+
 /**
  * Send a user message to the agent and start the agent loop.
  * Returns immediately; events are streamed via the event bus.
  */
-export async function runPrompt(sessionId, message, selectedCapabilities = [], promptCommand = null, transientContextPrompt = null) {
+export async function runPrompt(sessionId, message, selectedCapabilities = [], promptCommand = null, transientContextPrompt = null, modelAccessContext = null) {
   let session = await syncSessionFromStorage(sessionId)
   if (!session) {
     throw Object.assign(new Error('Session not found'), { statusCode: 404 })
   }
+  if (modelAccessContext) session.modelAccessContext = modelAccessContext
+  await refreshSessionModelBinding(session)
 
   if (session.agent.state.isStreaming || session.abortPending) {
-    throw Object.assign(new Error('Generation is still running. Stop it or wait until it finishes.'), { statusCode: 409 })
+    throw Object.assign(new Error('Generation is still running. Stop it or wait until it finishes.'), {
+      statusCode: 409,
+      errorCode: 'GENERATION_ALREADY_RUNNING',
+    })
   }
 
   await refreshMemoryState(session)
@@ -2149,7 +2236,7 @@ export async function runPrompt(sessionId, message, selectedCapabilities = [], p
     ? { role: 'user', content: message, timestamp: new Date().toISOString() }
     : message
   const commandState = await resolveCommandState(session, initialUserMessage, promptCommand)
-  const userMessage = commandState.userMessage ?? initialUserMessage
+  const resolvedUserMessage = commandState.userMessage ?? initialUserMessage
 
   if (commandState.textResponse) {
     session.agent.state.messages = [
@@ -2176,6 +2263,8 @@ export async function runPrompt(sessionId, message, selectedCapabilities = [], p
     return compactSession(session, initialUserMessage, commandState.compact)
   }
 
+  const userMessage = prepareCloudUserMessage(session, resolvedUserMessage)
+
   // Set a meaningful fallback immediately. The AI title request starts only
   // after the first user message has been persisted by the message_end handler.
   if (session.titleSource === 'default' && session.title === 'New chat') {
@@ -2221,14 +2310,19 @@ export async function runPrompt(sessionId, message, selectedCapabilities = [], p
  * Trims messages to keep up to and including the last user message,
  * removing the assistant response that follows it.
  */
-export async function continueSession(sessionId) {
+export async function continueSession(sessionId, modelAccessContext = null) {
   const session = agentSessions.get(sessionId)
   if (!session) {
     throw Object.assign(new Error('Session not found'), { statusCode: 404 })
   }
   if (session.agent.state.isStreaming) {
-    throw Object.assign(new Error('Generation is still running. Stop it or wait until it finishes.'), { statusCode: 409 })
+    throw Object.assign(new Error('Generation is still running. Stop it or wait until it finishes.'), {
+      statusCode: 409,
+      errorCode: 'GENERATION_ALREADY_RUNNING',
+    })
   }
+  if (modelAccessContext) session.modelAccessContext = modelAccessContext
+  await refreshSessionModelBinding(session)
 
   const messages = Array.isArray(session.agent.state.messages) ? session.agent.state.messages : []
 
@@ -2246,10 +2340,13 @@ export async function continueSession(sessionId) {
 
   const lastUserMessage = messages[lastUserIndex]
   const commandState = await resolveCommandState(session, lastUserMessage)
-  const continuedUserMessage = commandState.userMessage ?? lastUserMessage
+  const continuedUserMessage = prepareCloudUserMessage(session, commandState.userMessage ?? lastUserMessage)
   const trimmedMessages = messages.slice(0, lastUserIndex).concat(continuedUserMessage)
   updateSessionMessages(session, trimmedMessages)
   resetSessionCompaction(session)
+  if (isManagedCloudModel(session.model) && logicalMessageId(continuedUserMessage)) {
+    await flushSessionPersist(session)
+  }
 
   resetIdleTimer(session)
   session.activeCommandName = commandState.commandName ?? null
@@ -2334,9 +2431,9 @@ export function steerAgent(sessionId, message) {
     throw Object.assign(new Error('Session not found'), { statusCode: 404 })
   }
 
-  const agentMessage = typeof message === 'string'
+  const agentMessage = prepareCloudUserMessage(session, typeof message === 'string'
     ? { role: 'user', content: message, timestamp: Date.now() }
-    : message
+    : message)
 
   session.agent.steer(agentMessage)
   return { sessionId, steered: true }
@@ -2351,9 +2448,9 @@ export function followUpAgent(sessionId, message) {
     throw Object.assign(new Error('Session not found'), { statusCode: 404 })
   }
 
-  const agentMessage = typeof message === 'string'
+  const agentMessage = prepareCloudUserMessage(session, typeof message === 'string'
     ? { role: 'user', content: message, timestamp: Date.now() }
-    : message
+    : message)
 
   session.agent.followUp(agentMessage)
   return { sessionId, followUp: true }
@@ -2443,6 +2540,7 @@ export function getSessionState(sessionId) {
     yoloMode: session.yoloMode,
     systemPrompt: session.agent.state.systemPrompt,
     model: session.model,
+    modelRef: session.modelRef || undefined,
     thinkingLevel: session.thinkingLevel,
     title: session.title,
     titleSource: session.titleSource,
@@ -2597,6 +2695,8 @@ export async function restoreAgent(sessionId) {
       accessMode: normalizeAccessMode(sessionData.accessMode, sessionData.yoloMode),
       yoloMode: sessionData.yoloMode || false,
       model: sessionData.model,
+      modelRef: sessionData.modelRef || null,
+      resolvePersistedModel: true,
       thinkingLevel: sessionData.thinkingLevel || 'off',
       messages: sessionData.messages || [],
       title: sessionData.title || 'New chat',
@@ -2752,7 +2852,7 @@ export async function updateSessionYoloMode(sessionId, yoloMode) {
  * Does NOT force persistence — normal lifecycle events (message_end, agent_end) will persist
  * the updated model.
  */
-export function updateSessionModel(sessionId, model) {
+export function updateSessionModel(sessionId, model, modelRef = null) {
   const session = agentSessions.get(sessionId)
   if (!session) {
     throw Object.assign(new Error('Session not found'), { statusCode: 404 })
@@ -2762,9 +2862,11 @@ export function updateSessionModel(sessionId, model) {
   }
 
   session.model = model
+  session.modelRef = modelRef
+  session.resolvePersistedModel = true
   session.agent.state.model = model
 
-  return { sessionId, model }
+  return { sessionId, model, modelRef: modelRef || undefined }
 }
 
 /**

package/server/agent-profile-files.mjs

@@ -6,7 +6,7 @@ import { logger } from './utils/logger.mjs'
 import { firstOptionalBoolean, firstString, parseFrontmatter, splitDelimitedList } from './frontmatter.mjs'
 import { normalizeAgentProfileThinkingLevel, normalizeCapabilityPolicy, normalizeModelReference, validateAgentProfileTools, validateModelReference } from './agent-profile-schema.mjs'
 
-const DEFAULT_MAX_RUNTIME_MS = 30 * 60 * 1000
+const DEFAULT_MAX_RUNTIME_MS = 60 * 60 * 1000
 const DEFAULT_MAX_TOOL_CALLS = 300
 const nameRegex = /^[a-z][a-z0-9_-]{1,39}$/
 const allowedToolNames = new Set(['read_file', 'grep_files', 'write_file', 'edit_file', 'run_command'])
@@ -66,14 +66,18 @@ function quoteFrontmatter(value) {
 
 function formatModelFrontmatter(model) {
   if (!model || model.mode !== 'fixed') return 'model:\n  mode: inherit'
-  return [
+  const result = [
     'model:',
     '  mode: fixed',
-    `  provider: ${quoteFrontmatter(model.provider)}`,
-    `  modelId: ${quoteFrontmatter(model.modelId)}`,
+    model.source && model.source !== 'legacy-custom' ? `  source: ${quoteFrontmatter(model.source)}` : '',
+    model.providerId ? `  providerId: ${quoteFrontmatter(model.providerId)}` : '',
+    model.catalogId ? `  catalogId: ${quoteFrontmatter(model.catalogId)}` : '',
+    model.provider ? `  provider: ${quoteFrontmatter(model.provider)}` : '',
+    model.modelId ? `  modelId: ${quoteFrontmatter(model.modelId)}` : '',
     model.api ? `  api: ${quoteFrontmatter(model.api)}` : '',
     model.baseUrl ? `  baseUrl: ${quoteFrontmatter(model.baseUrl)}` : '',
   ].filter(Boolean).join('\n')
+  return result
 }
 
 export function userAgentProfileFilePath(name) {

package/server/agent-profile-schema.mjs

@@ -1,3 +1,5 @@
+import { resolveModelBinding } from './model-catalog.mjs'
+
 export const AGENT_PROFILE_TOOL_NAMES = ['read_file', 'grep_files', 'write_file', 'edit_file', 'run_command']
 export const AGENT_PROFILE_MODEL_INHERIT = { mode: 'inherit' }
 export const AGENT_PROFILE_THINKING_LEVELS = ['inherit', 'off', 'low', 'medium', 'high', 'xhigh']
@@ -41,10 +43,6 @@ function normalizeString(value) {
   return typeof value === 'string' && value.trim() ? value.trim() : undefined
 }
 
-function sameBaseUrl(a, b) {
-  return String(a || '').trim().replace(/\/$/, '') === String(b || '').trim().replace(/\/$/, '')
-}
-
 function readModelField(value, ...keys) {
   if (!value || typeof value !== 'object' || Array.isArray(value)) return undefined
   for (const key of keys) {
@@ -90,10 +88,16 @@ export function normalizeModelReference(input, metadata = {}) {
   const flatModelId = normalizeString(metadata['model-id'] ?? metadata.model_id ?? metadata.modelId)
   const flatApi = normalizeString(metadata['model-api'] ?? metadata.modelApi ?? metadata.api)
   const flatBaseUrl = normalizeString(metadata['model-base-url'] ?? metadata.modelBaseUrl ?? metadata.baseUrl)
+  const flatSource = normalizeString(metadata['model-source'] ?? metadata.modelSource)?.toLowerCase()
+  const flatProviderId = normalizeString(metadata['model-provider-id'] ?? metadata.modelProviderId ?? metadata.providerId)
+  const flatCatalogId = normalizeString(metadata['model-catalog-id'] ?? metadata.modelCatalogId ?? metadata.catalogId)
 
-  if (flatMode === 'fixed' || flatProvider || flatModelId) {
+  if (flatMode === 'fixed' || flatSource || flatProvider || flatModelId || flatProviderId || flatCatalogId) {
     return {
       mode: 'fixed',
+      source: flatSource,
+      providerId: flatProviderId,
+      catalogId: flatCatalogId,
       provider: flatProvider,
       modelId: flatModelId,
       api: flatApi,
@@ -123,6 +127,9 @@ export function normalizeModelReference(input, metadata = {}) {
 
   return {
     mode: 'fixed',
+    source: readModelField(modelValue, 'source'),
+    providerId: readModelField(modelValue, 'providerId', 'provider-id', 'provider_id'),
+    catalogId: readModelField(modelValue, 'catalogId', 'catalog-id', 'catalog_id', 'quickforgeCatalogId'),
     provider: readModelField(modelValue, 'provider'),
     modelId: readModelField(modelValue, 'modelId', 'model-id', 'model_id', 'id', 'model'),
     api: readModelField(modelValue, 'api'),
@@ -133,9 +140,19 @@ export function normalizeModelReference(input, metadata = {}) {
 export function validateModelReference(modelRef) {
   const ref = modelRef || AGENT_PROFILE_MODEL_INHERIT
   if (ref.mode !== 'fixed') return { ...AGENT_PROFILE_MODEL_INHERIT }
+  if (ref.source === 'cloud') {
+    const catalogId = ref.catalogId || ref.modelId
+    if (!catalogId) throw requestError('Fixed Cloud agent model requires catalogId')
+    return { mode: 'fixed', source: 'cloud', catalogId }
+  }
+  if (ref.source === 'custom') {
+    if (!ref.providerId || !ref.modelId) throw requestError('Fixed custom agent model requires providerId and modelId')
+    return { mode: 'fixed', source: 'custom', providerId: ref.providerId, modelId: ref.modelId }
+  }
   if (!ref.provider || !ref.modelId) throw requestError('Fixed agent model requires provider and modelId')
   return {
     mode: 'fixed',
+    source: 'legacy-custom',
     provider: ref.provider,
     modelId: ref.modelId,
     api: ref.api,
@@ -164,14 +181,17 @@ export function modelReferenceSnapshot(modelRef) {
   if (ref.mode !== 'fixed') return { ...AGENT_PROFILE_MODEL_INHERIT }
   return {
     mode: 'fixed',
-    provider: ref.provider,
-    modelId: ref.modelId,
+    ...(ref.source && ref.source !== 'legacy-custom' ? { source: ref.source } : {}),
+    ...(ref.providerId ? { providerId: ref.providerId } : {}),
+    ...(ref.catalogId ? { catalogId: ref.catalogId } : {}),
+    ...(ref.provider ? { provider: ref.provider } : {}),
+    ...(ref.modelId ? { modelId: ref.modelId } : {}),
     ...(ref.api ? { api: ref.api } : {}),
     ...(ref.baseUrl ? { baseUrl: ref.baseUrl } : {}),
   }
 }
 
-export async function resolveAgentProfileModel(profile, parentModel, readStore) {
+export async function resolveAgentProfileModel(profile, parentModel, _readStore, context = {}) {
   const ref = profile?.model || AGENT_PROFILE_MODEL_INHERIT
   if (!ref || ref.mode !== 'fixed') {
     if (!parentModel) throw requestError('No active model is configured for the parent session.')
@@ -179,19 +199,23 @@ export async function resolveAgentProfileModel(profile, parentModel, readStore)
   }
 
   const validRef = validateModelReference(ref)
-  const store = await readStore('custom-providers').catch(() => ({}))
-  const providers = Array.isArray(store) ? store : Object.values(store || {})
-  const configuredModels = providers
-    .flatMap((provider) => Array.isArray(provider?.models) ? provider.models : [])
-    .filter((model) => model && typeof model === 'object')
-  const model = configuredModels.find((candidate) => (
-    candidate.provider === validRef.provider &&
-    candidate.id === validRef.modelId &&
-    (!validRef.api || candidate.api === validRef.api) &&
-    (!validRef.baseUrl || sameBaseUrl(candidate.baseUrl, validRef.baseUrl))
-  ))
-  if (!model) {
-    throw requestError(`Agent profile ${profile?.name || ''} requested model ${validRef.provider}/${validRef.modelId}, but it is not configured.`)
-  }
-  return { model, info: { ...modelReferenceSnapshot(validRef), inherited: false } }
+  const modelRef = validRef.source === 'cloud'
+    ? { version: 1, source: 'cloud', catalogId: validRef.catalogId }
+    : validRef.source === 'custom'
+      ? { version: 1, source: 'custom', providerId: validRef.providerId, modelId: validRef.modelId }
+      : {
+          version: 1,
+          source: 'legacy-custom',
+          provider: validRef.provider,
+          modelId: validRef.modelId,
+          ...(validRef.api ? { api: validRef.api } : {}),
+          ...(validRef.baseUrl ? { baseUrl: validRef.baseUrl } : {}),
+        }
+  const binding = await resolveModelBinding({ modelRef }, {
+    context,
+    currentModel: parentModel,
+    allowCurrentHidden: true,
+    forExecution: true,
+  })
+  return { model: binding.model, modelRef: binding.modelRef, info: { ...modelReferenceSnapshot(validRef), inherited: false } }
 }

package/server/agent-profiles.mjs

@@ -27,7 +27,7 @@ const BUILTIN_OVERRIDES_STORE = 'agent-profile-overrides'
 const RESERVED_NAMES = new Set(subagentDefinitions.map((definition) => definition.name))
 const allowedToolNames = new Set(AGENT_PROFILE_TOOL_NAMES)
 const nameRegex = /^[a-z][a-z0-9_-]{1,39}$/
-const DEFAULT_MAX_RUNTIME_MS = 30 * 60 * 1000
+const DEFAULT_MAX_RUNTIME_MS = 60 * 60 * 1000
 const DEFAULT_MAX_TOOL_CALLS = 300
 const MIGRATION_MARKER_FILE = path.join(userAgentsDir, '.custom-agents-migrated.json')
 let customAgentMigrationPromise = null