From 47b19d4f1701a2a5be83ccb6e9742eb278880e6a Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Mon, 6 Apr 2026 23:11:29 +0800
Subject: [PATCH 01/15] docs: update README badges and project structure

Update website link to openalice.ai, add docs badge, and rewrite
Project Structure section to match current monorepo layout (packages/,
ui/, server/, removed plugins/ and skills/).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 README.md | 162 ++++++++++++++++++++++++++++--------------------------
 1 file changed, 85 insertions(+), 77 deletions(-)

diff --git a/README.md b/README.md
index 3f93fb8a..bb66fd77 100644
--- a/README.md
+++ b/README.md
@@ -3,7 +3,7 @@
 </p>
 
 <p align="center">
-  <a href="https://github.com/TraderAlice/OpenAlice/actions/workflows/ci.yml"><img src="https://github.com/TraderAlice/OpenAlice/actions/workflows/ci.yml/badge.svg" alt="CI"></a> · <a href="LICENSE"><img src="https://img.shields.io/badge/License-AGPL--3.0-blue.svg" alt="License: AGPL-3.0"></a> · <a href="https://deepwiki.com/TraderAlice/OpenAlice"><img src="https://deepwiki.com/badge.svg" alt="Ask DeepWiki"></a> · <a href="https://traderalice.com"><img src="https://img.shields.io/badge/Website-Visit-blue" alt="traderalice.com"></a>
+  <a href="https://github.com/TraderAlice/OpenAlice/actions/workflows/ci.yml"><img src="https://github.com/TraderAlice/OpenAlice/actions/workflows/ci.yml/badge.svg" alt="CI"></a> · <a href="LICENSE"><img src="https://img.shields.io/badge/License-AGPL--3.0-blue.svg" alt="License: AGPL-3.0"></a> · <a href="https://openalice.ai"><img src="https://img.shields.io/badge/Website-openalice.ai-blue" alt="openalice.ai"></a> · <a href="https://openalice.ai/docs"><img src="https://img.shields.io/badge/Docs-Read-green" alt="Docs"></a> · <a href="https://deepwiki.com/TraderAlice/OpenAlice"><img src="https://deepwiki.com/badge.svg" alt="Ask DeepWiki"></a>
 </p>
 
 # Open Alice
@@ -196,85 +196,93 @@ On first run, defaults are auto-copied to the user override path. Edit the user
 
 ## Project Structure
 
+Open Alice is a pnpm monorepo with Turborepo build orchestration.
+
 ```
+packages/
+├── ibkr/                      # @traderalice/ibkr — IBKR TWS API TypeScript port
+└── opentypebb/                # @traderalice/opentypebb — OpenBB platform TS port
+ui/                            # React frontend (Vite, 13 pages)
 src/
-  main.ts                    # Composition root — wires everything together
-  core/
-    agent-center.ts          # Top-level AI orchestration, owns ProviderRouter
-    ai-provider-manager.ts   # GenerateRouter + StreamableResult + AskOptions
-    tool-center.ts           # Centralized tool registry (Vercel + MCP export)
-    session.ts               # JSONL session store + format converters
-    compaction.ts            # Auto-summarize long context windows
-    config.ts                # Zod-validated config loader (generic account schema with brokerConfig)
-    ai-config.ts             # Runtime AI provider selection
-    event-log.ts             # Append-only JSONL event log
-    connector-center.ts      # ConnectorCenter — push delivery + last-interacted tracking
-    async-channel.ts         # AsyncChannel for streaming provider events to SSE
-    model-factory.ts         # Model instance factory for Vercel AI SDK
-    media.ts                 # MediaAttachment extraction
-    media-store.ts           # Media file persistence
-    types.ts                 # Plugin, EngineContext interfaces
-  ai-providers/
-    vercel-ai-sdk/           # Vercel AI SDK ToolLoopAgent wrapper
-    agent-sdk/               # Claude backend (@anthropic-ai/claude-agent-sdk, OAuth + API key)
-  domain/
-    trading/                 # Unified multi-account trading, guard pipeline, git-like commits
-      UnifiedTradingAccount.ts  # UTA class — owns broker + git + guards + snapshots
-      account-manager.ts     # UTA lifecycle (init, reconnect, enable/disable) + registry
-      git-persistence.ts     # Git state load/save
-      brokers/
-        registry.ts          # Broker self-registration (configSchema + configFields + fromConfig)
-        alpaca/              # Alpaca (US equities)
-        ccxt/                # CCXT (100+ crypto exchanges)
-        ibkr/                # Interactive Brokers (TWS/Gateway)
-        mock/                # In-memory test broker
-      git/                   # Trading-as-Git engine (stage → commit → push)
-      guards/                # Pre-execution safety checks (position size, cooldown, whitelist)
-      snapshot/              # Periodic + event-driven account state capture, equity curve
-    market-data/             # Structured data layer (typebb in-process + OpenBB API remote)
-      equity/                # Equity data + SymbolIndex (SEC/TMX local cache)
-      crypto/                # Crypto data layer
-      currency/              # Currency/forex data layer
-      commodity/             # Commodity data layer (EIA, spot prices)
-      economy/               # Macro economy data layer
-      client/                # Data backend clients (typebb SDK, openbb-api)
-    analysis/                # Indicators, technical analysis
-    news/                    # RSS collector + archive search
-    brain/                   # Cognitive state (memory, emotion)
-    thinking/                # Safe expression evaluator
-  tool/                      # AI tool definitions — thin bridge from domain to ToolCenter
-    trading.ts               # Trading tools (delegates to domain/trading)
-    equity.ts                # Equity fundamental tools (uses domain/market-data)
-    market.ts                # Symbol search tools (uses domain/market-data)
-    analysis.ts              # Indicator calculation tools (uses domain/analysis)
-    news.ts                  # News archive tools (uses domain/news)
-    brain.ts                 # Cognition tools (uses domain/brain)
-    thinking.ts              # Reasoning tools (uses domain/thinking)
-    browser.ts               # Browser automation tools (wraps openclaw)
-  server/
-    mcp.ts                   # MCP protocol server
-    opentypebb.ts            # Embedded OpenBB-compatible HTTP API (optional)
-  connectors/
-    web/                     # Web UI chat (Hono, SSE streaming, sub-channels)
-    telegram/                # Telegram bot (grammY, polling, commands)
-    mcp-ask/                 # MCP Ask connector (external agent conversation)
-  task/
-    cron/                    # Cron scheduling (engine, listener, AI tools)
-    heartbeat/               # Periodic heartbeat with structured response protocol
-  openclaw/                  # ⚠️ Frozen — DO NOT MODIFY
+├── main.ts                    # Composition root — wires everything together
+├── core/
+│   ├── agent-center.ts        # Top-level AI orchestration, owns ProviderRouter
+│   ├── ai-provider-manager.ts # GenerateRouter + StreamableResult + AskOptions
+│   ├── tool-center.ts         # Centralized tool registry (Vercel + MCP export)
+│   ├── mcp-export.ts          # Shared MCP export layer with type coercion
+│   ├── session.ts             # JSONL session store + format converters
+│   ├── compaction.ts          # Auto-summarize long context windows
+│   ├── config.ts              # Zod-validated config loader
+│   ├── event-log.ts           # Append-only JSONL event log
+│   ├── connector-center.ts    # ConnectorCenter — push delivery + last-interacted tracking
+│   ├── async-channel.ts       # AsyncChannel for streaming provider events to SSE
+│   ├── tool-call-log.ts       # Tool invocation logging
+│   ├── media.ts               # MediaAttachment extraction
+│   ├── media-store.ts         # Media file persistence
+│   └── types.ts               # Plugin, EngineContext interfaces
+├── ai-providers/
+│   ├── vercel-ai-sdk/         # Vercel AI SDK ToolLoopAgent wrapper
+│   ├── agent-sdk/             # Claude backend (@anthropic-ai/claude-agent-sdk, OAuth + API key)
+│   └── mock/                  # Mock provider (testing)
+├── domain/
+│   ├── trading/               # Unified multi-account trading, guard pipeline, git-like commits
+│   │   ├── account-manager.ts # UTA lifecycle (init, reconnect, enable/disable) + registry
+│   │   ├── git-persistence.ts # Git state load/save
+│   │   ├── brokers/
+│   │   │   ├── registry.ts    # Broker self-registration (configSchema + configFields + fromConfig)
+│   │   │   ├── alpaca/        # Alpaca (US equities)
+│   │   │   ├── ccxt/          # CCXT (100+ crypto exchanges)
+│   │   │   ├── ibkr/          # Interactive Brokers (TWS/Gateway)
+│   │   │   └── mock/          # In-memory test broker
+│   │   ├── git/               # Trading-as-Git engine (stage → commit → push)
+│   │   ├── guards/            # Pre-execution safety checks (position size, cooldown, whitelist)
+│   │   └── snapshot/          # Periodic + event-driven account state capture, equity curve
+│   ├── market-data/           # Structured data layer (opentypebb in-process + OpenBB API remote)
+│   │   ├── equity/            # Equity data + SymbolIndex (SEC/TMX local cache)
+│   │   ├── crypto/            # Crypto data layer
+│   │   ├── currency/          # Currency/forex data layer
+│   │   ├── commodity/         # Commodity data layer (EIA, spot prices)
+│   │   ├── economy/           # Macro economy data layer
+│   │   └── client/            # Data backend clients (opentypebb SDK, openbb-api)
+│   ├── analysis/              # Indicators, technical analysis
+│   ├── news/                  # RSS collector + archive search
+│   ├── brain/                 # Cognitive state (memory, emotion)
+│   └── thinking/              # Safe expression evaluator
+├── tool/                      # AI tool definitions — thin bridge from domain to ToolCenter
+│   ├── trading.ts             # Trading tools (delegates to domain/trading)
+│   ├── equity.ts              # Equity fundamental tools
+│   ├── market.ts              # Symbol search tools
+│   ├── analysis.ts            # Indicator calculation tools
+│   ├── news.ts                # News archive tools
+│   ├── brain.ts               # Cognition tools
+│   ├── thinking.ts            # Reasoning tools
+│   ├── browser.ts             # Browser automation tools (wraps openclaw)
+│   └── session.ts             # Session awareness tools
+├── server/
+│   ├── mcp.ts                 # MCP protocol server
+│   └── opentypebb.ts          # Embedded OpenBB-compatible HTTP API (optional)
+├── connectors/
+│   ├── web/                   # Web UI (Hono, SSE streaming, sub-channels)
+│   ├── telegram/              # Telegram bot (grammY, magic link auth, /trading panel)
+│   ├── mcp-ask/               # MCP Ask connector (external agent conversation)
+│   └── mock/                  # Mock connector (testing)
+├── task/
+│   ├── cron/                  # Cron scheduling (engine, listener, AI tools)
+│   └── heartbeat/             # Periodic heartbeat with structured response protocol
+└── openclaw/                  # ⚠️ Frozen — DO NOT MODIFY
 data/
-  config/                    # JSON configuration files
-  sessions/                  # JSONL conversation histories
-  brain/                     # Agent memory and emotion logs
-  cache/                     # API response caches
-  trading/                   # Trading commit history + snapshots (per-account)
-  news-collector/            # Persistent news archive (JSONL)
-  cron/                      # Cron job definitions (jobs.json)
-  event-log/                 # Persistent event log (events.jsonl)
-  tool-calls/                # Tool invocation logs
-  media/                     # Uploaded attachments
-default/                     # Factory defaults (persona, heartbeat prompts)
-docs/                        # Architecture documentation
+├── config/                    # JSON configuration files
+├── sessions/                  # JSONL conversation histories (web/, telegram/, cron/)
+├── brain/                     # Agent memory and emotion logs
+├── cache/                     # API response caches
+├── trading/                   # Trading commit history + snapshots (per-account)
+├── news-collector/            # Persistent news archive (JSONL)
+├── cron/                      # Cron job definitions (jobs.json)
+├── event-log/                 # Persistent event log (events.jsonl)
+├── tool-calls/                # Tool invocation logs
+└── media/                     # Uploaded attachments
+default/                       # Factory defaults (persona, heartbeat, skills)
+docs/                          # Documentation
 ```
 
 ## Roadmap to v1

From 0819e7c951c80b54983f09ec9d3d393274dbd910 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 10:52:28 +0800
Subject: [PATCH 02/15] fix: persona & heartbeat default file auto-copy + UI
 editing

- Add heartbeat auto-copy at startup (same as persona)
- Heartbeat GET route falls back to default/ when data file missing
- Add persona GET/PUT backend routes with default fallback
- Add persona editor to Settings page in frontend

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 src/connectors/web/routes/heartbeat.ts |  5 ++
 src/connectors/web/routes/persona.ts   | 40 +++++++++++++
 src/connectors/web/web-plugin.ts       |  2 +
 src/main.ts                            |  3 +
 ui/src/api/index.ts                    |  2 +
 ui/src/api/persona.ts                  | 18 ++++++
 ui/src/pages/SettingsPage.tsx          | 80 ++++++++++++++++++++++++++
 7 files changed, 150 insertions(+)
 create mode 100644 src/connectors/web/routes/persona.ts
 create mode 100644 ui/src/api/persona.ts

diff --git a/src/connectors/web/routes/heartbeat.ts b/src/connectors/web/routes/heartbeat.ts
index ec374d48..64090835 100644
--- a/src/connectors/web/routes/heartbeat.ts
+++ b/src/connectors/web/routes/heartbeat.ts
@@ -4,6 +4,7 @@ import { dirname } from 'node:path'
 import type { EngineContext } from '../../../core/types.js'
 
 const PROMPT_FILE = 'data/brain/heartbeat.md'
+const PROMPT_DEFAULT = 'default/heartbeat.default.md'
 
 /** Heartbeat routes: GET /status, POST /trigger, PUT /enabled, GET/PUT /prompt-file */
 export function createHeartbeatRoutes(ctx: EngineContext) {
@@ -43,6 +44,10 @@ export function createHeartbeatRoutes(ctx: EngineContext) {
       return c.json({ content, path: PROMPT_FILE })
     } catch (err: unknown) {
       if (err instanceof Error && 'code' in err && (err as NodeJS.ErrnoException).code === 'ENOENT') {
+        try {
+          const fallback = await readFile(PROMPT_DEFAULT, 'utf-8')
+          return c.json({ content: fallback, path: PROMPT_FILE })
+        } catch { /* default also missing */ }
         return c.json({ content: '', path: PROMPT_FILE })
       }
       return c.json({ error: String(err) }, 500)
diff --git a/src/connectors/web/routes/persona.ts b/src/connectors/web/routes/persona.ts
new file mode 100644
index 00000000..eb460829
--- /dev/null
+++ b/src/connectors/web/routes/persona.ts
@@ -0,0 +1,40 @@
+import { Hono } from 'hono'
+import { readFile, writeFile, mkdir } from 'node:fs/promises'
+import { dirname } from 'node:path'
+
+const PERSONA_FILE = 'data/brain/persona.md'
+const PERSONA_DEFAULT = 'default/persona.default.md'
+
+/** Persona routes: GET / (read), PUT / (write) */
+export function createPersonaRoutes() {
+  const app = new Hono()
+
+  app.get('/', async (c) => {
+    try {
+      const content = await readFile(PERSONA_FILE, 'utf-8')
+      return c.json({ content, path: PERSONA_FILE })
+    } catch (err: unknown) {
+      if (err instanceof Error && 'code' in err && (err as NodeJS.ErrnoException).code === 'ENOENT') {
+        try {
+          const fallback = await readFile(PERSONA_DEFAULT, 'utf-8')
+          return c.json({ content: fallback, path: PERSONA_FILE })
+        } catch { /* default also missing */ }
+        return c.json({ content: '', path: PERSONA_FILE })
+      }
+      return c.json({ error: String(err) }, 500)
+    }
+  })
+
+  app.put('/', async (c) => {
+    try {
+      const { content } = await c.req.json<{ content: string }>()
+      await mkdir(dirname(PERSONA_FILE), { recursive: true })
+      await writeFile(PERSONA_FILE, content, 'utf-8')
+      return c.json({ ok: true })
+    } catch (err) {
+      return c.json({ error: String(err) }, 500)
+    }
+  })
+
+  return app
+}
diff --git a/src/connectors/web/web-plugin.ts b/src/connectors/web/web-plugin.ts
index 4dbdcfdb..0243422e 100644
--- a/src/connectors/web/web-plugin.ts
+++ b/src/connectors/web/web-plugin.ts
@@ -18,6 +18,7 @@ import { createTradingConfigRoutes } from './routes/trading-config.js'
 import { createDevRoutes } from './routes/dev.js'
 import { createToolsRoutes } from './routes/tools.js'
 import { createAgentStatusRoutes } from './routes/agent-status.js'
+import { createPersonaRoutes } from './routes/persona.js'
 
 export interface WebConfig {
   port: number
@@ -83,6 +84,7 @@ export class WebPlugin implements Plugin {
     app.route('/api/dev', createDevRoutes(ctx.connectorCenter))
     app.route('/api/tools', createToolsRoutes(ctx.toolCenter))
     app.route('/api/agent-status', createAgentStatusRoutes(ctx))
+    app.route('/api/persona', createPersonaRoutes())
 
     // ==================== Serve UI (Vite build output) ====================
     const uiRoot = resolve('dist/ui')
diff --git a/src/main.ts b/src/main.ts
index 85d170e2..0f05c3c0 100644
--- a/src/main.ts
+++ b/src/main.ts
@@ -49,6 +49,8 @@ const FRONTAL_LOBE_FILE = resolve('data/brain/frontal-lobe.md')
 const EMOTION_LOG_FILE = resolve('data/brain/emotion-log.md')
 const PERSONA_FILE = resolve('data/brain/persona.md')
 const PERSONA_DEFAULT = resolve('default/persona.default.md')
+const HEARTBEAT_FILE = resolve('data/brain/heartbeat.md')
+const HEARTBEAT_DEFAULT = resolve('default/heartbeat.default.md')
 
 const sleep = (ms: number) => new Promise<void>((r) => setTimeout(r, ms))
 
@@ -99,6 +101,7 @@ async function main() {
   const [brainExport, persona] = await Promise.all([
     readFile(BRAIN_FILE, 'utf-8').then((r) => JSON.parse(r) as BrainExportState).catch(() => undefined),
     readWithDefault(PERSONA_FILE, PERSONA_DEFAULT),
+    readWithDefault(HEARTBEAT_FILE, HEARTBEAT_DEFAULT),
   ])
 
   const brainDir = resolve('data/brain')
diff --git a/ui/src/api/index.ts b/ui/src/api/index.ts
index 84fcdbad..94fed442 100644
--- a/ui/src/api/index.ts
+++ b/ui/src/api/index.ts
@@ -13,6 +13,7 @@ import { devApi } from './dev'
 import { toolsApi } from './tools'
 import { channelsApi } from './channels'
 import { agentStatusApi } from './agentStatus'
+import { personaApi } from './persona'
 export const api = {
   chat: chatApi,
   config: configApi,
@@ -25,6 +26,7 @@ export const api = {
   tools: toolsApi,
   channels: channelsApi,
   agentStatus: agentStatusApi,
+  persona: personaApi,
 }
 
 // Re-export all types for convenience
diff --git a/ui/src/api/persona.ts b/ui/src/api/persona.ts
new file mode 100644
index 00000000..0e665664
--- /dev/null
+++ b/ui/src/api/persona.ts
@@ -0,0 +1,18 @@
+import { headers } from './client'
+
+export const personaApi = {
+  async get(): Promise<{ content: string; path: string }> {
+    const res = await fetch('/api/persona')
+    if (!res.ok) throw new Error('Failed to load persona')
+    return res.json()
+  },
+
+  async update(content: string): Promise<void> {
+    const res = await fetch('/api/persona', {
+      method: 'PUT',
+      headers,
+      body: JSON.stringify({ content }),
+    })
+    if (!res.ok) throw new Error('Failed to save persona')
+  },
+}
diff --git a/ui/src/pages/SettingsPage.tsx b/ui/src/pages/SettingsPage.tsx
index f3f43861..cea18ca1 100644
--- a/ui/src/pages/SettingsPage.tsx
+++ b/ui/src/pages/SettingsPage.tsx
@@ -48,6 +48,11 @@ export function SettingsPage() {
               </div>
             </ConfigSection>
 
+            {/* Persona */}
+            <ConfigSection title="Persona" description="The system prompt that defines Alice's personality and behavior. Changes take effect on next server restart.">
+              <PersonaEditor />
+            </ConfigSection>
+
             {/* Compaction */}
             <ConfigSection title="Compaction" description="Context window management. When conversation size approaches Max Context minus Max Output tokens, older messages are automatically summarized to free up space.">
               <CompactionForm config={config} />
@@ -91,3 +96,78 @@ function CompactionForm({ config }: { config: AppConfig }) {
   )
 }
 
+// ==================== Persona Editor ====================
+
+function PersonaEditor() {
+  const [content, setContent] = useState('')
+  const [filePath, setFilePath] = useState('')
+  const [loading, setLoading] = useState(true)
+  const [saving, setSaving] = useState(false)
+  const [saved, setSaved] = useState(false)
+  const [error, setError] = useState<string | null>(null)
+  const [dirty, setDirty] = useState(false)
+
+  useEffect(() => {
+    api.persona.get()
+      .then(({ content, path }) => {
+        setContent(content)
+        setFilePath(path)
+      })
+      .catch(() => setError('Failed to load persona'))
+      .finally(() => setLoading(false))
+  }, [])
+
+  const handleSave = async () => {
+    setSaving(true)
+    setError(null)
+    setSaved(false)
+    try {
+      await api.persona.update(content)
+      setDirty(false)
+      setSaved(true)
+      setTimeout(() => setSaved(false), 2000)
+    } catch {
+      setError('Failed to save')
+    } finally {
+      setSaving(false)
+    }
+  }
+
+  if (loading) return <div className="text-sm text-text-muted">Loading...</div>
+
+  return (
+    <>
+      <textarea
+        className={`${inputClass} min-h-[200px] max-h-[400px] resize-y font-mono text-xs leading-relaxed`}
+        value={content}
+        onChange={(e) => { setContent(e.target.value); setDirty(true) }}
+      />
+      <div className="flex items-center gap-2 mt-2">
+        <button
+          onClick={handleSave}
+          disabled={saving || !dirty}
+          className="btn-primary-sm"
+        >
+          {saving ? 'Saving...' : 'Save'}
+        </button>
+        {saved && (
+          <span className="inline-flex items-center gap-1.5 text-[11px]">
+            <span className="w-1.5 h-1.5 rounded-full bg-green" />
+            <span className="text-text-muted">Saved</span>
+          </span>
+        )}
+        {error && (
+          <span className="inline-flex items-center gap-1.5 text-[11px]">
+            <span className="w-1.5 h-1.5 rounded-full bg-red" />
+            <span className="text-red">{error}</span>
+          </span>
+        )}
+        {dirty && !saved && !error && (
+          <span className="text-[11px] text-text-muted">Unsaved changes</span>
+        )}
+      </div>
+      {filePath && <p className="text-[11px] text-text-muted mt-1">{filePath}</p>}
+    </>
+  )
+}
+

From 43877902288c79079d2e49a0dc4c418130588ad4 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 11:14:38 +0800
Subject: [PATCH 03/15] feat: hot-reload persona and brain state in system
 prompt

Instructions are now rebuilt on each request instead of being frozen at
startup. Persona is re-read from disk and brain state (frontal lobe,
emotion) is pulled live from the Brain instance. Both providers accept
a () => Promise<string> getter; VercelAIProvider cache invalidates when
instructions content changes.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 .../agent-sdk/agent-sdk-provider.ts           |  5 +--
 .../vercel-ai-sdk/vercel-provider.spec.ts     |  6 ++--
 .../vercel-ai-sdk/vercel-provider.ts          | 11 ++++---
 src/core/agent-center.spec.ts                 |  2 +-
 src/main.ts                                   | 32 +++++++++++--------
 5 files changed, 32 insertions(+), 24 deletions(-)

diff --git a/src/ai-providers/agent-sdk/agent-sdk-provider.ts b/src/ai-providers/agent-sdk/agent-sdk-provider.ts
index ec50fcbb..f5991ca1 100644
--- a/src/ai-providers/agent-sdk/agent-sdk-provider.ts
+++ b/src/ai-providers/agent-sdk/agent-sdk-provider.ts
@@ -25,7 +25,7 @@ export class AgentSdkProvider implements AIProvider {
 
   constructor(
     private getTools: () => Promise<Record<string, Tool>>,
-    private systemPrompt?: string,
+    private getSystemPrompt: () => Promise<string>,
   ) {}
 
   /** Re-read agent config from disk to pick up hot-reloaded settings. */
@@ -46,6 +46,7 @@ export class AgentSdkProvider implements AIProvider {
 
   async ask(prompt: string): Promise<ProviderResult> {
     const config = await this.resolveConfig()
+    config.systemPrompt = await this.getSystemPrompt()
     const mcpServer = await this.buildMcpServer()
     const result = await askAgentSdk(prompt, config, undefined, mcpServer)
     return { text: result.text, media: [] }
@@ -62,7 +63,7 @@ export class AgentSdkProvider implements AIProvider {
       ...(opts?.disabledTools?.length
         ? { disallowedTools: [...(config.disallowedTools ?? []), ...opts.disabledTools] }
         : {}),
-      systemPrompt: opts?.systemPrompt ?? this.systemPrompt,
+      systemPrompt: opts?.systemPrompt ?? await this.getSystemPrompt(),
     }
 
     const override: AgentSdkOverride | undefined = opts?.agentSdk
diff --git a/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts b/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts
index 22f2af56..4fed6838 100644
--- a/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts
+++ b/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts
@@ -29,7 +29,7 @@ function makeAgent(text = 'ok', steps: any[] = []) {
 
 function makeProvider(overrides?: { getTools?: () => Promise<Record<string, any>> }) {
   const getTools = overrides?.getTools ?? (async () => ({ toolA: {}, toolB: {} }))
-  return new VercelAIProvider(getTools as any, 'You are a trading assistant.', 10)
+  return new VercelAIProvider(getTools as any, async () => 'You are a trading assistant.', 10)
 }
 
 // ==================== resolveAgent caching ====================
@@ -65,7 +65,7 @@ describe('VercelAIProvider — agent caching', () => {
 
   it('recreates agent when tool count changes', async () => {
     let toolSet: Record<string, any> = { toolA: {}, toolB: {} }
-    const provider = new VercelAIProvider(async () => toolSet as any, 'prompt', 5)
+    const provider = new VercelAIProvider(async () => toolSet as any, async () => 'prompt', 5)
     await provider.ask('first')
     toolSet = { toolA: {}, toolB: {}, toolC: {} }
     await provider.ask('second')
@@ -84,7 +84,7 @@ describe('VercelAIProvider — per-request overrides', () => {
 
   it('skips cache and uses filtered tools when disabledTools provided', async () => {
     const getTools = async () => ({ toolA: {} as any, toolB: {} as any, toolC: {} as any })
-    const provider = new VercelAIProvider(getTools, 'prompt', 5)
+    const provider = new VercelAIProvider(getTools, async () => 'prompt', 5)
     // warm cache
     await provider.ask('warm')
     vi.clearAllMocks()
diff --git a/src/ai-providers/vercel-ai-sdk/vercel-provider.ts b/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
index a4ba59ea..7fa3b47f 100644
--- a/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
+++ b/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
@@ -22,11 +22,12 @@ export class VercelAIProvider implements AIProvider {
   private cachedKey: string | null = null
   private cachedToolCount: number = 0
   private cachedSystemPrompt: string | null = null
+  private cachedInstructions: string | null = null
   private cachedAgent: Agent | null = null
 
   constructor(
     private getTools: () => Promise<Record<string, Tool>>,
-    private instructions: string,
+    private getInstructions: () => Promise<string>,
     private maxSteps: number,
   ) {}
 
@@ -34,6 +35,7 @@ export class VercelAIProvider implements AIProvider {
   private async resolveAgent(systemPrompt?: string, disabledTools?: string[], modelOverride?: ModelOverride): Promise<Agent> {
     const { model, key } = await createModelFromConfig(modelOverride)
     const allTools = await this.getTools()
+    const instructions = await this.getInstructions()
 
     // Per-channel overrides: skip cache and create a fresh agent
     if (disabledTools?.length || modelOverride) {
@@ -41,16 +43,17 @@ export class VercelAIProvider implements AIProvider {
       const tools = disabledSet
         ? Object.fromEntries(Object.entries(allTools).filter(([name]) => !disabledSet.has(name)))
         : allTools
-      return createAgent(model, tools, systemPrompt ?? this.instructions, this.maxSteps)
+      return createAgent(model, tools, systemPrompt ?? instructions, this.maxSteps)
     }
 
     const toolCount = Object.keys(allTools).length
     const effectivePrompt = systemPrompt ?? null
-    if (key !== this.cachedKey || toolCount !== this.cachedToolCount || effectivePrompt !== this.cachedSystemPrompt) {
-      this.cachedAgent = createAgent(model, allTools, systemPrompt ?? this.instructions, this.maxSteps)
+    if (key !== this.cachedKey || toolCount !== this.cachedToolCount || effectivePrompt !== this.cachedSystemPrompt || instructions !== this.cachedInstructions) {
+      this.cachedAgent = createAgent(model, allTools, systemPrompt ?? instructions, this.maxSteps)
       this.cachedKey = key
       this.cachedToolCount = toolCount
       this.cachedSystemPrompt = effectivePrompt
+      this.cachedInstructions = instructions
       console.log(`vercel-ai: model loaded → ${key} (${toolCount} tools)`)
     }
     return this.cachedAgent!
diff --git a/src/core/agent-center.spec.ts b/src/core/agent-center.spec.ts
index 0da85eec..61f41abd 100644
--- a/src/core/agent-center.spec.ts
+++ b/src/core/agent-center.spec.ts
@@ -43,7 +43,7 @@ function makeAgentCenter(overrides: MakeAgentCenterOpts = {}): AgentCenter {
   const compaction = overrides.compaction ?? DEFAULT_COMPACTION_CONFIG
 
   vi.mocked(createModelFromConfig).mockResolvedValue({ model, key: 'test:mock-model' })
-  const provider = new VercelAIProvider(async () => tools, instructions, maxSteps)
+  const provider = new VercelAIProvider(async () => tools, async () => instructions, maxSteps)
   const router = new GenerateRouter(provider, null)
 
   return new AgentCenter({ router, compaction })
diff --git a/src/main.ts b/src/main.ts
index 0f05c3c0..a4d03941 100644
--- a/src/main.ts
+++ b/src/main.ts
@@ -98,7 +98,7 @@ async function main() {
 
   // ==================== Brain ====================
 
-  const [brainExport, persona] = await Promise.all([
+  const [brainExport] = await Promise.all([
     readFile(BRAIN_FILE, 'utf-8').then((r) => JSON.parse(r) as BrainExportState).catch(() => undefined),
     readWithDefault(PERSONA_FILE, PERSONA_DEFAULT),
     readWithDefault(HEARTBEAT_FILE, HEARTBEAT_DEFAULT),
@@ -123,17 +123,21 @@ async function main() {
     ? Brain.restore(brainExport, { onCommit: brainOnCommit })
     : new Brain({ onCommit: brainOnCommit })
 
-  const frontalLobe = brain.getFrontalLobe()
-  const emotion = brain.getEmotion().current
-  const instructions = [
-    persona,
-    '---',
-    '## Current Brain State',
-    '',
-    `**Frontal Lobe:** ${frontalLobe || '(empty)'}`,
-    '',
-    `**Emotion:** ${emotion}`,
-  ].join('\n')
+  /** Re-read persona from disk + live brain state on each request. */
+  const getInstructions = async () => {
+    const persona = await readFile(PERSONA_FILE, 'utf-8').catch(() => '')
+    const frontalLobe = brain.getFrontalLobe()
+    const emotion = brain.getEmotion().current
+    return [
+      persona,
+      '---',
+      '## Current Brain State',
+      '',
+      `**Frontal Lobe:** ${frontalLobe || '(empty)'}`,
+      '',
+      `**Emotion:** ${emotion}`,
+    ].join('\n')
+  }
 
   // ==================== Cron ====================
 
@@ -212,12 +216,12 @@ async function main() {
 
   const vercelProvider = new VercelAIProvider(
     () => toolCenter.getVercelTools(),
-    instructions,
+    getInstructions,
     config.agent.maxSteps,
   )
   const agentSdkProvider = new AgentSdkProvider(
     () => toolCenter.getVercelTools(),
-    instructions,
+    getInstructions,
   )
   const router = new GenerateRouter(vercelProvider, agentSdkProvider)
 

From f704c4173270956549207bfdb4c78afc8f27d1a8 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 11:46:35 +0800
Subject: [PATCH 04/15] fix: remove stale createAgent re-export + fix spec type
 errors

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 src/ai-providers/vercel-ai-sdk/index.ts       |   2 -
 .../vercel-ai-sdk/vercel-provider.spec.ts     | 114 ++++++++----------
 2 files changed, 51 insertions(+), 65 deletions(-)

diff --git a/src/ai-providers/vercel-ai-sdk/index.ts b/src/ai-providers/vercel-ai-sdk/index.ts
index f981f7f9..dd3df74d 100644
--- a/src/ai-providers/vercel-ai-sdk/index.ts
+++ b/src/ai-providers/vercel-ai-sdk/index.ts
@@ -1,3 +1 @@
-export { createAgent } from './agent.js'
-export type { Agent } from './agent.js'
 export { VercelAIProvider } from './vercel-provider.js'
diff --git a/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts b/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts
index 4fed6838..0d3b24b8 100644
--- a/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts
+++ b/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts
@@ -6,7 +6,8 @@ vi.mock('./model-factory.js', () => ({
 }))
 
 vi.mock('./agent.js', () => ({
-  createAgent: vi.fn(),
+  generateText: vi.fn(),
+  stepCountIs: vi.fn().mockReturnValue(() => false),
 }))
 
 vi.mock('../../core/media.js', () => ({
@@ -14,125 +15,101 @@ vi.mock('../../core/media.js', () => ({
 }))
 
 import { createModelFromConfig } from './model-factory.js'
-import { createAgent } from './agent.js'
+import { generateText } from './agent.js'
 
 const mockCreateModelFromConfig = vi.mocked(createModelFromConfig)
-const mockCreateAgent = vi.mocked(createAgent)
+const mockGenerateText = vi.mocked(generateText)
 
 // ==================== Helpers ====================
 
-function makeAgent(text = 'ok', steps: any[] = []) {
-  return {
-    generate: vi.fn().mockResolvedValue({ text, steps }),
-  }
-}
-
 function makeProvider(overrides?: { getTools?: () => Promise<Record<string, any>> }) {
   const getTools = overrides?.getTools ?? (async () => ({ toolA: {}, toolB: {} }))
   return new VercelAIProvider(getTools as any, async () => 'You are a trading assistant.', 10)
 }
 
-// ==================== resolveAgent caching ====================
+// ==================== ask() ====================
 
-describe('VercelAIProvider — agent caching', () => {
+describe('VercelAIProvider — ask()', () => {
   beforeEach(() => {
     vi.clearAllMocks()
     mockCreateModelFromConfig.mockResolvedValue({ model: {} as any, key: 'gpt-4o' })
-    mockCreateAgent.mockReturnValue(makeAgent() as any)
+    mockGenerateText.mockResolvedValue({ text: 'ok', steps: [] } as any)
   })
 
-  it('creates agent on first ask()', async () => {
+  it('calls generateText with model, tools, system, and prompt', async () => {
     const provider = makeProvider()
     await provider.ask('hello')
-    expect(mockCreateAgent).toHaveBeenCalledOnce()
-  })
 
-  it('reuses cached agent on second ask() when nothing changes', async () => {
-    const provider = makeProvider()
-    await provider.ask('first')
-    await provider.ask('second')
-    expect(mockCreateAgent).toHaveBeenCalledOnce()
+    expect(mockGenerateText).toHaveBeenCalledOnce()
+    const call = mockGenerateText.mock.calls[0][0]
+    expect(call).toHaveProperty('prompt', 'hello')
+    expect(call).toHaveProperty('system', 'You are a trading assistant.')
+    expect(call.tools).toHaveProperty('toolA')
+    expect(call.tools).toHaveProperty('toolB')
   })
 
-  it('recreates agent when config key changes', async () => {
+  it('returns text from generateText result', async () => {
+    mockGenerateText.mockResolvedValue({ text: 'the answer', steps: [] } as any)
     const provider = makeProvider()
-    await provider.ask('first')
-    // Simulate config key change
-    mockCreateModelFromConfig.mockResolvedValue({ model: {} as any, key: 'claude-3-5-sonnet' })
-    await provider.ask('second')
-    expect(mockCreateAgent).toHaveBeenCalledTimes(2)
+    const result = await provider.ask('question')
+    expect(result.text).toBe('the answer')
   })
 
-  it('recreates agent when tool count changes', async () => {
-    let toolSet: Record<string, any> = { toolA: {}, toolB: {} }
-    const provider = new VercelAIProvider(async () => toolSet as any, async () => 'prompt', 5)
-    await provider.ask('first')
-    toolSet = { toolA: {}, toolB: {}, toolC: {} }
-    await provider.ask('second')
-    expect(mockCreateAgent).toHaveBeenCalledTimes(2)
+  it('returns empty string when text is null', async () => {
+    mockGenerateText.mockResolvedValue({ text: null, steps: [] } as any)
+    const provider = makeProvider()
+    const result = await provider.ask('question')
+    expect(result.text).toBe('')
   })
 })
 
-// ==================== per-request overrides ====================
+// ==================== generate() — tool filtering ====================
 
-describe('VercelAIProvider — per-request overrides', () => {
+describe('VercelAIProvider — generate() tool filtering', () => {
   beforeEach(() => {
     vi.clearAllMocks()
     mockCreateModelFromConfig.mockResolvedValue({ model: {} as any, key: 'gpt-4o' })
-    mockCreateAgent.mockReturnValue(makeAgent() as any)
+    mockGenerateText.mockResolvedValue({ text: 'ok', steps: [] } as any)
   })
 
-  it('skips cache and uses filtered tools when disabledTools provided', async () => {
+  it('filters disabled tools from generateText call', async () => {
     const getTools = async () => ({ toolA: {} as any, toolB: {} as any, toolC: {} as any })
     const provider = new VercelAIProvider(getTools, async () => 'prompt', 5)
-    // warm cache
-    await provider.ask('warm')
-    vi.clearAllMocks()
-    mockCreateModelFromConfig.mockResolvedValue({ model: {} as any, key: 'gpt-4o' })
-    mockCreateAgent.mockReturnValue(makeAgent() as any)
 
-    // generate with disabledTools
     const events = []
     for await (const e of provider.generate([], 'test', { disabledTools: ['toolB'] })) {
       events.push(e)
     }
 
-    expect(mockCreateAgent).toHaveBeenCalledOnce()
-    // The tools passed to createAgent should exclude toolB
-    const toolsArg = mockCreateAgent.mock.calls[0][1]
-    expect(Object.keys(toolsArg)).toContain('toolA')
-    expect(Object.keys(toolsArg)).not.toContain('toolB')
-    expect(Object.keys(toolsArg)).toContain('toolC')
+    const call = mockGenerateText.mock.calls[0][0]
+    const toolNames = Object.keys(call.tools!)
+    expect(toolNames).toContain('toolA')
+    expect(toolNames).not.toContain('toolB')
+    expect(toolNames).toContain('toolC')
   })
 
-  it('skips cache when modelOverride is provided', async () => {
+  it('passes modelOverride to createModelFromConfig', async () => {
     const provider = makeProvider()
-    await provider.ask('warm')
-    vi.clearAllMocks()
-    mockCreateModelFromConfig.mockResolvedValue({ model: {} as any, key: 'gpt-4o' })
-    mockCreateAgent.mockReturnValue(makeAgent() as any)
 
     for await (const _ of provider.generate([], 'test', { vercelAiSdk: { modelId: 'claude-3-7' } as any })) {
       // drain
     }
 
-    expect(mockCreateAgent).toHaveBeenCalledOnce()
+    expect(mockCreateModelFromConfig).toHaveBeenCalledWith({ modelId: 'claude-3-7' })
   })
 })
 
-// ==================== generate() behavior ====================
+// ==================== generate() — events ====================
 
-describe('VercelAIProvider — generate()', () => {
+describe('VercelAIProvider — generate() events', () => {
   beforeEach(() => {
     vi.clearAllMocks()
     mockCreateModelFromConfig.mockResolvedValue({ model: {} as any, key: 'gpt-4o' })
-    mockCreateAgent.mockReturnValue(makeAgent() as any)
   })
 
   it('yields done event with text from result', async () => {
+    mockGenerateText.mockResolvedValue({ text: 'final answer', steps: [] } as any)
     const provider = makeProvider()
-    const agent = makeAgent('final answer')
-    mockCreateAgent.mockReturnValue(agent as any)
 
     const events = []
     for await (const e of provider.generate([], 'test')) {
@@ -143,9 +120,8 @@ describe('VercelAIProvider — generate()', () => {
     expect(done?.result.text).toBe('final answer')
   })
 
-  it('propagates agent error through channel', async () => {
-    const agent = { generate: vi.fn().mockRejectedValue(new Error('model error')) }
-    mockCreateAgent.mockReturnValue(agent as any)
+  it('propagates error through channel', async () => {
+    mockGenerateText.mockRejectedValue(new Error('model error'))
     const provider = makeProvider()
 
     await expect(async () => {
@@ -154,4 +130,16 @@ describe('VercelAIProvider — generate()', () => {
       }
     }).rejects.toThrow('model error')
   })
+
+  it('uses per-channel systemPrompt override', async () => {
+    mockGenerateText.mockResolvedValue({ text: 'ok', steps: [] } as any)
+    const provider = makeProvider()
+
+    for await (const _ of provider.generate([], 'test', { systemPrompt: 'custom prompt' })) {
+      // drain
+    }
+
+    const call = mockGenerateText.mock.calls[0][0]
+    expect(call.system).toBe('custom prompt')
+  })
 })

From 6e1ecbe67878529326e1856d4670e9ef7a37445b Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 11:46:46 +0800
Subject: [PATCH 05/15] refactor: replace ToolLoopAgent with direct
 generateText calls

ToolLoopAgent was a thin wrapper around generateText with no meaningful
extras. Using generateText directly removes the agent instance cache
and simplifies the code. System prompt is now passed per-request.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 src/ai-providers/vercel-ai-sdk/agent.ts       | 27 ++-----
 .../vercel-ai-sdk/vercel-provider.ts          | 70 ++++++++-----------
 2 files changed, 36 insertions(+), 61 deletions(-)

diff --git a/src/ai-providers/vercel-ai-sdk/agent.ts b/src/ai-providers/vercel-ai-sdk/agent.ts
index 5528477d..4bbe0365 100644
--- a/src/ai-providers/vercel-ai-sdk/agent.ts
+++ b/src/ai-providers/vercel-ai-sdk/agent.ts
@@ -1,24 +1,9 @@
-import { ToolLoopAgent, stepCountIs } from 'ai'
-import type { LanguageModel, Tool } from 'ai'
-
 /**
- * Create a generic ToolLoopAgent with externally-provided tools.
+ * Re-export generateText as the primary tool-loop entry point.
  *
- * The caller decides what tools the agent has — Engine wires in
- * sandbox-analysis tools (market data, trading, cognition, etc.).
+ * Previously wrapped Vercel AI SDK's ToolLoopAgent, but that was a thin
+ * wrapper around generateText with no meaningful extras. Using generateText
+ * directly is simpler and avoids caching agent instances.
  */
-export function createAgent(
-  model: LanguageModel,
-  tools: Record<string, Tool>,
-  instructions: string,
-  maxSteps = 20,
-) {
-  return new ToolLoopAgent({
-    model,
-    tools,
-    instructions,
-    stopWhen: stepCountIs(maxSteps),
-  })
-}
-
-export type Agent = ReturnType<typeof createAgent>
+export { generateText, stepCountIs } from 'ai'
+export type { LanguageModel, Tool, StepResult } from 'ai'
diff --git a/src/ai-providers/vercel-ai-sdk/vercel-provider.ts b/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
index 7fa3b47f..eb76a8fa 100644
--- a/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
+++ b/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
@@ -1,29 +1,23 @@
 /**
- * VercelAIProvider — GenerateProvider backed by Vercel AI SDK's ToolLoopAgent.
+ * VercelAIProvider — GenerateProvider backed by Vercel AI SDK's generateText.
  *
- * The model is lazily created from config and cached.  When model.json or
- * api-keys.json changes on disk, the next request picks up the new model
- * automatically (hot-reload).
+ * The model is lazily created from config on each call. When ai-provider.json
+ * changes on disk, the next request picks up the new model automatically.
+ * Instructions (persona + brain state) are also re-read per request.
  */
 
 import type { ModelMessage, Tool } from 'ai'
 import type { ProviderResult, ProviderEvent, AIProvider, GenerateOpts } from '../types.js'
 import type { SessionEntry } from '../../core/session.js'
-import type { Agent } from './agent.js'
 import type { MediaAttachment } from '../../core/types.js'
 import { toModelMessages } from '../../core/session.js'
 import { extractMediaFromToolOutput } from '../../core/media.js'
 import { createModelFromConfig, type ModelOverride } from './model-factory.js'
-import { createAgent } from './agent.js'
+import { generateText, stepCountIs } from './agent.js'
 import { createChannel } from '../../core/async-channel.js'
 
 export class VercelAIProvider implements AIProvider {
   readonly providerTag = 'vercel-ai' as const
-  private cachedKey: string | null = null
-  private cachedToolCount: number = 0
-  private cachedSystemPrompt: string | null = null
-  private cachedInstructions: string | null = null
-  private cachedAgent: Agent | null = null
 
   constructor(
     private getTools: () => Promise<Record<string, Tool>>,
@@ -31,58 +25,54 @@ export class VercelAIProvider implements AIProvider {
     private maxSteps: number,
   ) {}
 
-  /** Lazily create or return the cached agent, re-creating when config, tools, or system prompt change. */
-  private async resolveAgent(systemPrompt?: string, disabledTools?: string[], modelOverride?: ModelOverride): Promise<Agent> {
-    const { model, key } = await createModelFromConfig(modelOverride)
-    const allTools = await this.getTools()
-    const instructions = await this.getInstructions()
+  /** Resolve model, tools, and instructions for a single request. */
+  private async resolve(disabledTools?: string[], modelOverride?: ModelOverride) {
+    const [{ model }, allTools, instructions] = await Promise.all([
+      createModelFromConfig(modelOverride),
+      this.getTools(),
+      this.getInstructions(),
+    ])
 
-    // Per-channel overrides: skip cache and create a fresh agent
-    if (disabledTools?.length || modelOverride) {
-      const disabledSet = disabledTools?.length ? new Set(disabledTools) : null
-      const tools = disabledSet
-        ? Object.fromEntries(Object.entries(allTools).filter(([name]) => !disabledSet.has(name)))
-        : allTools
-      return createAgent(model, tools, systemPrompt ?? instructions, this.maxSteps)
-    }
+    const tools = disabledTools?.length
+      ? Object.fromEntries(Object.entries(allTools).filter(([name]) => !new Set(disabledTools).has(name)))
+      : allTools
 
-    const toolCount = Object.keys(allTools).length
-    const effectivePrompt = systemPrompt ?? null
-    if (key !== this.cachedKey || toolCount !== this.cachedToolCount || effectivePrompt !== this.cachedSystemPrompt || instructions !== this.cachedInstructions) {
-      this.cachedAgent = createAgent(model, allTools, systemPrompt ?? instructions, this.maxSteps)
-      this.cachedKey = key
-      this.cachedToolCount = toolCount
-      this.cachedSystemPrompt = effectivePrompt
-      this.cachedInstructions = instructions
-      console.log(`vercel-ai: model loaded → ${key} (${toolCount} tools)`)
-    }
-    return this.cachedAgent!
+    return { model, tools, instructions }
   }
 
   async ask(prompt: string): Promise<ProviderResult> {
-    const agent = await this.resolveAgent(undefined)
+    const { model, tools, instructions } = await this.resolve()
     const media: MediaAttachment[] = []
-    const result = await agent.generate({
+
+    const result = await generateText({
+      model,
+      tools,
+      system: instructions,
       prompt,
+      stopWhen: stepCountIs(this.maxSteps),
       onStepFinish: (step) => {
         for (const tr of step.toolResults) {
           media.push(...extractMediaFromToolOutput(tr.output))
         }
       },
     })
+
     return { text: result.text ?? '', media }
   }
 
   async *generate(entries: SessionEntry[], _prompt: string, opts?: GenerateOpts): AsyncGenerator<ProviderEvent> {
+    const { model, tools, instructions } = await this.resolve(opts?.disabledTools, opts?.vercelAiSdk)
     const messages = toModelMessages(entries)
 
-    const agent = await this.resolveAgent(opts?.systemPrompt, opts?.disabledTools, opts?.vercelAiSdk)
-
     const channel = createChannel<ProviderEvent>()
     const media: MediaAttachment[] = []
 
-    const resultPromise = agent.generate({
+    const resultPromise = generateText({
+      model,
+      tools,
+      system: opts?.systemPrompt ?? instructions,
       messages: messages as ModelMessage[],
+      stopWhen: stepCountIs(this.maxSteps),
       onStepFinish: (step) => {
         for (const tc of step.toolCalls) {
           channel.push({ type: 'tool_use', id: tc.toolCallId, name: tc.toolName, input: tc.input })

From 0c413b6d3ec2d709b5a2f7d7797a63df1f68d371 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 11:59:52 +0800
Subject: [PATCH 06/15] fix: resolve remaining tsc errors in bbProvider specs
 and MockBroker

Type-assert executor.execute() return as unknown[] in bbProvider
integration tests. Remove invalid reduceOnly from MockBroker
closePosition (not part of TpSlParams interface).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 .../market-data/__tests__/bbProviders/fmp.bbProvider.spec.ts    | 2 +-
 .../__tests__/bbProviders/yfinance.bbProvider.spec.ts           | 2 +-
 src/domain/trading/brokers/mock/MockBroker.ts                   | 2 +-
 3 files changed, 3 insertions(+), 3 deletions(-)

diff --git a/src/domain/market-data/__tests__/bbProviders/fmp.bbProvider.spec.ts b/src/domain/market-data/__tests__/bbProviders/fmp.bbProvider.spec.ts
index a93d300a..692993a0 100644
--- a/src/domain/market-data/__tests__/bbProviders/fmp.bbProvider.spec.ts
+++ b/src/domain/market-data/__tests__/bbProviders/fmp.bbProvider.spec.ts
@@ -13,7 +13,7 @@ let ctx: TestContext
 beforeAll(async () => { ctx = await getTestContext() })
 
 const exec = (model: string, params: Record<string, unknown> = {}) =>
-  ctx.executor.execute('fmp', model, params, ctx.credentials)
+  ctx.executor.execute('fmp', model, params, ctx.credentials) as Promise<unknown[]>
 
 describe('fmp — equity', () => {
   beforeEach(({ skip }) => { if (!hasCredential(ctx.credentials, 'fmp')) skip('no fmp_api_key') })
diff --git a/src/domain/market-data/__tests__/bbProviders/yfinance.bbProvider.spec.ts b/src/domain/market-data/__tests__/bbProviders/yfinance.bbProvider.spec.ts
index a4bdaadf..55413f1d 100644
--- a/src/domain/market-data/__tests__/bbProviders/yfinance.bbProvider.spec.ts
+++ b/src/domain/market-data/__tests__/bbProviders/yfinance.bbProvider.spec.ts
@@ -13,7 +13,7 @@ let ctx: TestContext
 beforeAll(async () => { ctx = await getTestContext() })
 
 const exec = (model: string, params: Record<string, unknown> = {}) =>
-  ctx.executor.execute('yfinance', model, params, ctx.credentials)
+  ctx.executor.execute('yfinance', model, params, ctx.credentials) as Promise<unknown[]>
 
 describe('yfinance — equity', () => {
   it('EquityQuote', async () => { expect((await exec('EquityQuote', { symbol: 'AAPL' })).length).toBeGreaterThan(0) })
diff --git a/src/domain/trading/brokers/mock/MockBroker.ts b/src/domain/trading/brokers/mock/MockBroker.ts
index acf47478..9026c222 100644
--- a/src/domain/trading/brokers/mock/MockBroker.ts
+++ b/src/domain/trading/brokers/mock/MockBroker.ts
@@ -318,7 +318,7 @@ export class MockBroker implements IBroker {
     order.orderType = 'MKT'
     order.totalQuantity = quantity ?? pos.quantity
 
-    return this.placeOrder(pos.contract, order, { reduceOnly: true })
+    return this.placeOrder(pos.contract, order)
   }
 
   // ---- Queries ----

From 04996cb39a1cb2cf05e462e778542d35d7dac1b1 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 19:15:36 +0800
Subject: [PATCH 07/15] =?UTF-8?q?feat:=20add=20Codex=20AI=20provider=20?=
 =?UTF-8?q?=E2=80=94=20OpenAI=20models=20via=20ChatGPT=20subscription=20OA?=
 =?UTF-8?q?uth?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Adds a new AI provider backend that calls OpenAI Codex models (gpt-5.x,
codex-mini-latest) through the ChatGPT subscription endpoint, using OAuth
tokens from ~/.codex/auth.json (created by `codex login`).

- New provider at src/ai-providers/codex/ with auth, tool-bridge, and
  manual tool loop over the Responses API
- Direct HTTP via `openai` SDK — no CLI subprocess, no Vercel wrapper
- Context managed by us (toTextHistory), tools injected per-request
- Token auto-refresh against auth.openai.com/oauth/token
- Replaces @openai/codex-sdk dependency with openai SDK

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 package.json                               |   1 +
 pnpm-lock.yaml                             |  20 ++
 src/ai-providers/codex/auth.ts             | 208 ++++++++++++++++++
 src/ai-providers/codex/codex-provider.ts   | 237 +++++++++++++++++++++
 src/ai-providers/codex/index.ts            |   1 +
 src/ai-providers/codex/tool-bridge.ts      |  51 +++++
 src/ai-providers/mock/index.ts             |   4 +-
 src/ai-providers/types.ts                  |   3 +-
 src/connectors/telegram/telegram-plugin.ts |   1 +
 src/connectors/web/routes/chat.ts          |   1 +
 src/core/agent-center.ts                   |   1 +
 src/core/ai-provider-manager.ts            |  13 +-
 src/core/config.ts                         |  13 +-
 src/core/session.ts                        |   2 +-
 src/main.ts                                |   7 +-
 15 files changed, 554 insertions(+), 9 deletions(-)
 create mode 100644 src/ai-providers/codex/auth.ts
 create mode 100644 src/ai-providers/codex/codex-provider.ts
 create mode 100644 src/ai-providers/codex/index.ts
 create mode 100644 src/ai-providers/codex/tool-bridge.ts

diff --git a/package.json b/package.json
index 5b24a59a..676fb8db 100644
--- a/package.json
+++ b/package.json
@@ -51,6 +51,7 @@
     "grammy": "^1.40.0",
     "hono": "^4.12.7",
     "json5": "^2.2.3",
+    "openai": "^6.33.0",
     "pino": "^10.3.1",
     "playwright-core": "1.58.2",
     "sharp": "^0.34.5",
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 62877d53..4d061825 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -75,6 +75,9 @@ importers:
       json5:
         specifier: ^2.2.3
         version: 2.2.3
+      openai:
+        specifier: ^6.33.0
+        version: 6.33.0(ws@8.19.0)(zod@4.3.6)
       pino:
         specifier: ^10.3.1
         version: 10.3.1
@@ -2376,6 +2379,18 @@ packages:
   once@1.4.0:
     resolution: {integrity: sha512-lNaJgI+2Q5URQBkccEKHTQOPaXdUxnZZElQTZY0MFUAuaEqe1E+Nyvgdz/aIyNi6Z9MzO5dv1H8n58/GELp3+w==}
 
+  openai@6.33.0:
+    resolution: {integrity: sha512-xAYN1W3YsDXJWA5F277135YfkEk6H7D3D6vWwRhJ3OEkzRgcyK8z/P5P9Gyi/wB4N8kK9kM5ZjprfvyHagKmpw==}
+    hasBin: true
+    peerDependencies:
+      ws: ^8.18.0
+      zod: ^3.25 || ^4.0
+    peerDependenciesMeta:
+      ws:
+        optional: true
+      zod:
+        optional: true
+
   parse5@8.0.0:
     resolution: {integrity: sha512-9m4m5GSgXjL4AjumKzq1Fgfp3Z8rsvjRNbnkVwfu2ImRqE5D0LnY2QfDen18FSY9C573YU5XxSapdHZTZ2WolA==}
 
@@ -5030,6 +5045,11 @@ snapshots:
     dependencies:
       wrappy: 1.0.2
 
+  openai@6.33.0(ws@8.19.0)(zod@4.3.6):
+    optionalDependencies:
+      ws: 8.19.0
+      zod: 4.3.6
+
   parse5@8.0.0:
     dependencies:
       entities: 6.0.1
diff --git a/src/ai-providers/codex/auth.ts b/src/ai-providers/codex/auth.ts
new file mode 100644
index 00000000..975c3a79
--- /dev/null
+++ b/src/ai-providers/codex/auth.ts
@@ -0,0 +1,208 @@
+/**
+ * Codex OAuth authentication — reads ~/.codex/auth.json and manages token refresh.
+ *
+ * Users authenticate via `codex login` (OpenAI Codex CLI). This module reads
+ * the cached OAuth tokens and refreshes them when expired, writing updates back
+ * to disk so the Codex CLI stays in sync.
+ */
+
+import { readFile, writeFile, mkdir } from 'node:fs/promises'
+import { join, dirname } from 'node:path'
+import { homedir } from 'node:os'
+import { pino } from 'pino'
+
+const logger = pino({
+  transport: { target: 'pino/file', options: { destination: 'logs/codex.log', mkdir: true } },
+})
+
+// ==================== Constants ====================
+
+const REFRESH_URL = 'https://auth.openai.com/oauth/token'
+const CLIENT_ID = 'app_EMoamEEZ73f0CkXaXp7hrann'
+
+// ==================== Types ====================
+
+export interface CodexAuthFile {
+  auth_mode?: string
+  OPENAI_API_KEY?: string | null
+  tokens?: {
+    id_token: string
+    access_token: string
+    refresh_token: string
+    account_id?: string
+  }
+  last_refresh?: string
+}
+
+// ==================== Helpers ====================
+
+/** Resolve the Codex home directory ($CODEX_HOME or ~/.codex). */
+export function resolveCodexHome(): string {
+  const env = process.env.CODEX_HOME
+  if (env) return env
+  return join(homedir(), '.codex')
+}
+
+function authFilePath(): string {
+  return join(resolveCodexHome(), 'auth.json')
+}
+
+/**
+ * Decode a JWT payload without signature verification.
+ * Returns the parsed claims object.
+ */
+function decodeJwtPayload(token: string): Record<string, unknown> {
+  const parts = token.split('.')
+  if (parts.length !== 3) throw new Error('Invalid JWT format')
+  return JSON.parse(Buffer.from(parts[1]!, 'base64url').toString())
+}
+
+/** Extract the `exp` (expiration) claim from a JWT as epoch seconds. */
+function getJwtExpiration(token: string): number | null {
+  try {
+    const claims = decodeJwtPayload(token)
+    return typeof claims.exp === 'number' ? claims.exp : null
+  } catch {
+    return null
+  }
+}
+
+// ==================== Core ====================
+
+/** Read and parse the auth.json file. */
+export async function readAuthFile(): Promise<CodexAuthFile> {
+  const path = authFilePath()
+  try {
+    const raw = await readFile(path, 'utf-8')
+    return JSON.parse(raw)
+  } catch (err: any) {
+    if (err?.code === 'ENOENT') {
+      throw new Error(
+        `Codex auth not found at ${path}. Run \`codex login\` to authenticate.`,
+      )
+    }
+    throw new Error(`Failed to read Codex auth file: ${err?.message}`)
+  }
+}
+
+/** Write the auth file back to disk (0o600 permissions on unix). */
+async function writeAuthFile(auth: CodexAuthFile): Promise<void> {
+  const path = authFilePath()
+  await mkdir(dirname(path), { recursive: true })
+  const json = JSON.stringify(auth, null, 2)
+  await writeFile(path, json, { mode: 0o600 })
+}
+
+/** Check whether the access token is expired or stale. */
+function isTokenExpired(auth: CodexAuthFile): boolean {
+  const token = auth.tokens?.access_token
+  if (!token) return true
+
+  // Check JWT exp claim
+  const exp = getJwtExpiration(token)
+  if (exp != null && exp <= Date.now() / 1000) return true
+
+  // Proactive refresh: 8+ days since last refresh
+  if (auth.last_refresh) {
+    const lastRefresh = new Date(auth.last_refresh).getTime()
+    const eightDays = 8 * 24 * 60 * 60 * 1000
+    if (Date.now() - lastRefresh > eightDays) return true
+  }
+
+  return false
+}
+
+/** Request new tokens from the OpenAI auth service. */
+async function refreshTokens(
+  refreshToken: string,
+): Promise<{ id_token?: string; access_token?: string; refresh_token?: string }> {
+  const res = await fetch(REFRESH_URL, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      client_id: CLIENT_ID,
+      grant_type: 'refresh_token',
+      refresh_token: refreshToken,
+    }),
+  })
+
+  if (!res.ok) {
+    const body = await res.text().catch(() => '')
+    logger.error({ status: res.status, body: body.slice(0, 500) }, 'token_refresh_failed')
+    throw new Error(
+      `Codex token refresh failed (${res.status}). You may need to run \`codex login\` again.`,
+    )
+  }
+
+  return await res.json() as { id_token?: string; access_token?: string; refresh_token?: string }
+}
+
+// ==================== In-memory cache ====================
+
+let cachedToken: { token: string; expiresAt: number } | null = null
+let refreshPromise: Promise<string> | null = null
+
+/**
+ * Get a valid access token, refreshing if necessary.
+ *
+ * Uses an in-memory cache to avoid repeated disk reads. A promise-based
+ * mutex prevents concurrent refreshes from racing on disk writes.
+ */
+export async function getAccessToken(): Promise<string> {
+  // Fast path: cached and not expired
+  if (cachedToken && cachedToken.expiresAt > Date.now() / 1000) {
+    return cachedToken.token
+  }
+
+  // Mutex: if a refresh is already in progress, wait for it
+  if (refreshPromise) return refreshPromise
+
+  refreshPromise = (async () => {
+    try {
+      const auth = await readAuthFile()
+
+      if (!auth.tokens?.access_token) {
+        throw new Error('Codex auth.json has no tokens. Run `codex login` to authenticate.')
+      }
+
+      if (!isTokenExpired(auth)) {
+        // Token is still valid — cache it
+        const exp = getJwtExpiration(auth.tokens.access_token)
+        cachedToken = {
+          token: auth.tokens.access_token,
+          expiresAt: exp ?? Date.now() / 1000 + 3600,
+        }
+        return cachedToken.token
+      }
+
+      // Token expired — refresh
+      logger.info('refreshing_codex_token')
+      const refreshed = await refreshTokens(auth.tokens.refresh_token)
+
+      // Merge refreshed tokens
+      if (refreshed.access_token) auth.tokens.access_token = refreshed.access_token
+      if (refreshed.refresh_token) auth.tokens.refresh_token = refreshed.refresh_token
+      if (refreshed.id_token) auth.tokens.id_token = refreshed.id_token
+      auth.last_refresh = new Date().toISOString()
+
+      await writeAuthFile(auth)
+      logger.info('codex_token_refreshed')
+
+      const exp = getJwtExpiration(auth.tokens.access_token)
+      cachedToken = {
+        token: auth.tokens.access_token,
+        expiresAt: exp ?? Date.now() / 1000 + 3600,
+      }
+      return cachedToken.token
+    } finally {
+      refreshPromise = null
+    }
+  })()
+
+  return refreshPromise
+}
+
+/** Clear the in-memory token cache (useful after auth errors). */
+export function clearTokenCache(): void {
+  cachedToken = null
+}
diff --git a/src/ai-providers/codex/codex-provider.ts b/src/ai-providers/codex/codex-provider.ts
new file mode 100644
index 00000000..0125aad2
--- /dev/null
+++ b/src/ai-providers/codex/codex-provider.ts
@@ -0,0 +1,237 @@
+/**
+ * CodexProvider — AIProvider backed by OpenAI Codex models via ChatGPT subscription OAuth.
+ *
+ * Calls the Responses API at chatgpt.com/backend-api/codex/responses using
+ * the standard OpenAI TypeScript SDK. Auth tokens are read from ~/.codex/auth.json
+ * (created by `codex login`).
+ *
+ * Context is managed by us — each call starts fresh (no previous_response_id).
+ * Tools are injected via the Responses API `tools` field.
+ */
+
+import OpenAI from 'openai'
+import type { Tool } from 'ai'
+import { pino } from 'pino'
+
+import type { ProviderResult, ProviderEvent, AIProvider, GenerateOpts } from '../types.js'
+import type { SessionEntry } from '../../core/session.js'
+import { toTextHistory } from '../../core/session.js'
+import { buildChatHistoryPrompt, DEFAULT_MAX_HISTORY } from '../utils.js'
+import { readAIProviderConfig, readAgentConfig } from '../../core/config.js'
+import { getAccessToken, clearTokenCache } from './auth.js'
+import { convertTools } from './tool-bridge.js'
+
+const logger = pino({
+  transport: { target: 'pino/file', options: { destination: 'logs/codex.log', mkdir: true } },
+})
+
+const DEFAULT_BASE_URL = 'https://chatgpt.com/backend-api/codex'
+const DEFAULT_MODEL = 'codex-mini-latest'
+
+// ==================== Provider ====================
+
+export class CodexProvider implements AIProvider {
+  readonly providerTag = 'codex' as const
+
+  constructor(
+    private getTools: () => Promise<Record<string, Tool>>,
+    private getSystemPrompt: () => Promise<string>,
+  ) {}
+
+  /** Create an OpenAI client with the current access token. */
+  private async createClient(opts?: GenerateOpts): Promise<{ client: OpenAI; model: string }> {
+    const token = await getAccessToken()
+    const aiConfig = await readAIProviderConfig()
+    const baseURL = opts?.codex?.baseUrl ?? DEFAULT_BASE_URL
+    const model = opts?.codex?.model ?? aiConfig.model ?? DEFAULT_MODEL
+
+    const client = new OpenAI({
+      apiKey: token,
+      baseURL,
+    })
+
+    return { client, model }
+  }
+
+  async ask(prompt: string): Promise<ProviderResult> {
+    const { client, model } = await this.createClient()
+    const instructions = await this.getSystemPrompt()
+
+    try {
+      const response = await client.responses.create({
+        model,
+        instructions,
+        input: prompt,
+      })
+
+      const text = response.output
+        .filter((item): item is OpenAI.Responses.ResponseOutputMessage => item.type === 'message')
+        .flatMap(msg => msg.content)
+        .filter((c): c is OpenAI.Responses.ResponseOutputText => c.type === 'output_text')
+        .map(c => c.text)
+        .join('')
+
+      return { text: text || '(no output)', media: [] }
+    } catch (err) {
+      logger.error({ err }, 'ask_error')
+      throw err
+    }
+  }
+
+  async *generate(
+    entries: SessionEntry[],
+    prompt: string,
+    opts?: GenerateOpts,
+  ): AsyncGenerator<ProviderEvent> {
+    const maxHistory = opts?.maxHistoryEntries ?? DEFAULT_MAX_HISTORY
+    const textHistory = toTextHistory(entries).slice(-maxHistory)
+    const fullPrompt = buildChatHistoryPrompt(prompt, textHistory, opts?.historyPreamble)
+
+    const { client, model } = await this.createClient(opts)
+    const instructions = opts?.systemPrompt ?? await this.getSystemPrompt()
+    const agentConfig = await readAgentConfig()
+    const maxSteps = agentConfig.maxSteps
+
+    // Build tools
+    const allTools = await this.getTools()
+    const tools = convertTools(allTools, opts?.disabledTools)
+
+    // Build initial input
+    const input: OpenAI.Responses.ResponseInputItem[] = [
+      { role: 'user', content: fullPrompt, type: 'message' },
+    ]
+
+    yield* this.toolLoop(client, model, instructions, input, tools, allTools, maxSteps)
+  }
+
+  /**
+   * The manual tool loop — sends requests to the Responses API and executes
+   * function calls until the model responds with text only or we hit maxSteps.
+   */
+  private async *toolLoop(
+    client: OpenAI,
+    model: string,
+    instructions: string,
+    input: OpenAI.Responses.ResponseInputItem[],
+    tools: ReturnType<typeof convertTools>,
+    vercelTools: Record<string, Tool>,
+    maxSteps: number,
+  ): AsyncGenerator<ProviderEvent> {
+    let accumulatedText = ''
+
+    for (let step = 0; step < maxSteps; step++) {
+      const functionCalls: Array<{
+        call_id: string
+        name: string
+        arguments: string
+      }> = []
+      let stepText = ''
+
+      try {
+        const stream = client.responses.stream({
+          model,
+          instructions,
+          input,
+          tools: tools.length > 0 ? tools : undefined,
+        })
+
+        for await (const event of stream) {
+          if (event.type === 'response.output_text.delta') {
+            yield { type: 'text', text: event.delta }
+            stepText += event.delta
+          } else if (event.type === 'response.function_call_arguments.done') {
+            functionCalls.push({
+              call_id: event.item_id,
+              name: event.name,
+              arguments: event.arguments,
+            })
+          }
+        }
+      } catch (err: any) {
+        // On 401, clear token cache and surface auth error
+        if (err?.status === 401) {
+          clearTokenCache()
+          const errorText = accumulatedText + stepText +
+            '\n\n[Codex auth expired. Run `codex login` to re-authenticate.]'
+          yield { type: 'done', result: { text: errorText, media: [] } }
+          return
+        }
+        logger.error({ err: err?.message, status: err?.status }, 'responses_api_error')
+        const errorText = accumulatedText + stepText +
+          `\n\n[Codex API error: ${err?.message ?? 'unknown error'}]`
+        yield { type: 'done', result: { text: errorText, media: [] } }
+        return
+      }
+
+      accumulatedText += stepText
+
+      // No function calls — model is done
+      if (functionCalls.length === 0) {
+        yield { type: 'done', result: { text: accumulatedText, media: [] } }
+        return
+      }
+
+      // Execute function calls and build follow-up input
+      const toolResults: Array<{ call_id: string; output: string }> = []
+
+      for (const fc of functionCalls) {
+        let parsedInput: unknown
+        try {
+          parsedInput = JSON.parse(fc.arguments)
+        } catch {
+          parsedInput = {}
+        }
+
+        // Yield tool_use event
+        yield { type: 'tool_use', id: fc.call_id, name: fc.name, input: parsedInput }
+        logger.info({ tool: fc.name, call_id: fc.call_id }, 'tool_use')
+
+        // Execute the tool
+        const tool = vercelTools[fc.name]
+        let resultContent: string
+        if (!tool?.execute) {
+          resultContent = JSON.stringify({ error: `Unknown tool: ${fc.name}` })
+        } else {
+          try {
+            const result = await tool.execute(parsedInput, {
+              toolCallId: fc.call_id,
+              messages: [],
+            })
+            resultContent = typeof result === 'string' ? result : JSON.stringify(result ?? '')
+          } catch (err) {
+            resultContent = JSON.stringify({ error: `Tool execution failed: ${err}` })
+          }
+        }
+
+        // Yield tool_result event
+        yield { type: 'tool_result', tool_use_id: fc.call_id, content: resultContent }
+        logger.info({ tool: fc.name, call_id: fc.call_id, content: resultContent.slice(0, 300) }, 'tool_result')
+
+        toolResults.push({ call_id: fc.call_id, output: resultContent })
+      }
+
+      // Append function calls + outputs to input for next round
+      for (const fc of functionCalls) {
+        input.push({
+          type: 'function_call',
+          call_id: fc.call_id,
+          name: fc.name,
+          arguments: fc.arguments,
+        } as OpenAI.Responses.ResponseInputItem)
+      }
+      for (const tr of toolResults) {
+        input.push({
+          type: 'function_call_output',
+          call_id: tr.call_id,
+          output: tr.output,
+        } as OpenAI.Responses.ResponseInputItem)
+      }
+    }
+
+    // Max steps reached
+    yield {
+      type: 'done',
+      result: { text: accumulatedText + '\n\n[Max tool iterations reached]', media: [] },
+    }
+  }
+}
diff --git a/src/ai-providers/codex/index.ts b/src/ai-providers/codex/index.ts
new file mode 100644
index 00000000..f051eb5e
--- /dev/null
+++ b/src/ai-providers/codex/index.ts
@@ -0,0 +1 @@
+export { CodexProvider } from './codex-provider.js'
diff --git a/src/ai-providers/codex/tool-bridge.ts b/src/ai-providers/codex/tool-bridge.ts
new file mode 100644
index 00000000..e1a727da
--- /dev/null
+++ b/src/ai-providers/codex/tool-bridge.ts
@@ -0,0 +1,51 @@
+/**
+ * Tool bridge — converts ToolCenter's Vercel AI SDK tools to OpenAI Responses API format.
+ *
+ * Much simpler than the Agent SDK bridge (no MCP server needed) — just JSON Schema objects.
+ */
+
+import { z } from 'zod'
+import type { Tool } from 'ai'
+
+// ==================== Types ====================
+
+export interface ResponsesApiTool {
+  type: 'function'
+  name: string
+  description: string
+  parameters: Record<string, unknown>
+  strict: boolean | null
+}
+
+// ==================== Conversion ====================
+
+/**
+ * Convert Vercel AI SDK tools to OpenAI Responses API tool definitions.
+ *
+ * @param tools  Record<name, Tool> from ToolCenter.getVercelTools()
+ * @param disabledTools  Optional list of tool names to exclude
+ */
+export function convertTools(
+  tools: Record<string, Tool>,
+  disabledTools?: string[],
+): ResponsesApiTool[] {
+  const disabledSet = new Set(disabledTools ?? [])
+
+  return Object.entries(tools)
+    .filter(([name, t]) => t.execute && !disabledSet.has(name))
+    .map(([name, t]) => {
+      let parameters: Record<string, unknown>
+      try {
+        parameters = z.toJSONSchema(t.inputSchema as z.ZodType)
+      } catch {
+        parameters = { type: 'object', properties: {} }
+      }
+      return {
+        type: 'function' as const,
+        name,
+        description: t.description ?? name,
+        parameters,
+        strict: null,
+      }
+    })
+}
diff --git a/src/ai-providers/mock/index.ts b/src/ai-providers/mock/index.ts
index 6a09a762..38c59c33 100644
--- a/src/ai-providers/mock/index.ts
+++ b/src/ai-providers/mock/index.ts
@@ -33,7 +33,7 @@ export interface MockAIProviderCall {
 // ==================== Options ====================
 
 export interface MockAIProviderOpts {
-  providerTag?: 'vercel-ai' | 'claude-code' | 'agent-sdk'
+  providerTag?: 'vercel-ai' | 'claude-code' | 'agent-sdk' | 'codex'
   /** Text returned by ask(). Default: 'mock-ask-result'. */
   askResult?: string
 }
@@ -41,7 +41,7 @@ export interface MockAIProviderOpts {
 // ==================== MockAIProvider ====================
 
 export class MockAIProvider implements AIProvider {
-  readonly providerTag: 'vercel-ai' | 'claude-code' | 'agent-sdk'
+  readonly providerTag: 'vercel-ai' | 'claude-code' | 'agent-sdk' | 'codex'
   readonly generateCalls: MockAIProviderCall[] = []
   readonly askCalls: string[] = []
   private _askResult: string
diff --git a/src/ai-providers/types.ts b/src/ai-providers/types.ts
index 0dd007df..b31ece4a 100644
--- a/src/ai-providers/types.ts
+++ b/src/ai-providers/types.ts
@@ -32,6 +32,7 @@ export interface GenerateOpts {
   disabledTools?: string[]
   vercelAiSdk?: { provider: string; model: string; baseUrl?: string; apiKey?: string }
   agentSdk?: { model?: string; apiKey?: string; baseUrl?: string }
+  codex?: { model?: string; baseUrl?: string }
 }
 
 // ==================== AIProvider ====================
@@ -44,7 +45,7 @@ export interface GenerateOpts {
  */
 export interface AIProvider {
   /** Session log provenance tag. */
-  readonly providerTag: 'vercel-ai' | 'claude-code' | 'agent-sdk'
+  readonly providerTag: 'vercel-ai' | 'claude-code' | 'agent-sdk' | 'codex'
   /** Stateless one-shot prompt (used for compaction summarization, etc.). */
   ask(prompt: string): Promise<ProviderResult>
   /** Stream events from the backend. Yields tool_use/tool_result/text, then done. */
diff --git a/src/connectors/telegram/telegram-plugin.ts b/src/connectors/telegram/telegram-plugin.ts
index 808a6700..a6499b90 100644
--- a/src/connectors/telegram/telegram-plugin.ts
+++ b/src/connectors/telegram/telegram-plugin.ts
@@ -21,6 +21,7 @@ const BACKEND_LABELS: Record<AIBackend, string> = {
   'claude-code': 'Claude Code',
   'vercel-ai-sdk': 'Vercel AI SDK',
   'agent-sdk': 'Agent SDK',
+  'codex': 'Codex',
 }
 
 export class TelegramPlugin implements Plugin {
diff --git a/src/connectors/web/routes/chat.ts b/src/connectors/web/routes/chat.ts
index 30bbf35d..354894ba 100644
--- a/src/connectors/web/routes/chat.ts
+++ b/src/connectors/web/routes/chat.ts
@@ -46,6 +46,7 @@ export function createChatRoutes({ ctx, sessions, sseByChannel }: ChatDeps) {
         if (channel.provider) opts.provider = channel.provider
         if (channel.vercelAiSdk) opts.vercelAiSdk = channel.vercelAiSdk
         if (channel.agentSdk) opts.agentSdk = channel.agentSdk
+        if (channel.codex) opts.codex = channel.codex
       }
     }
 
diff --git a/src/core/agent-center.ts b/src/core/agent-center.ts
index 2a3bf598..5c6afc6f 100644
--- a/src/core/agent-center.ts
+++ b/src/core/agent-center.ts
@@ -96,6 +96,7 @@ export class AgentCenter {
       disabledTools: opts?.disabledTools,
       vercelAiSdk: opts?.vercelAiSdk,
       agentSdk: opts?.agentSdk,
+      codex: opts?.codex,
     }
     const source = provider.generate(entries, prompt, genOpts)
 
diff --git a/src/core/ai-provider-manager.ts b/src/core/ai-provider-manager.ts
index 7b036e9e..65a615e0 100644
--- a/src/core/ai-provider-manager.ts
+++ b/src/core/ai-provider-manager.ts
@@ -109,7 +109,7 @@ export interface AskOptions {
    * AI provider to use for this call, overriding the global ai-provider-manager.json config.
    * Falls back to global config if not specified.
    */
-  provider?: 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk'
+  provider?: 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk' | 'codex'
   /**
    * Vercel AI SDK model override — per-request provider/model/baseUrl/apiKey.
    * Only used when the active backend is 'vercel-ai-sdk'.
@@ -129,6 +129,14 @@ export interface AskOptions {
     apiKey?: string
     baseUrl?: string
   }
+  /**
+   * Codex model override — per-request model/baseUrl.
+   * Only used when the active backend is 'codex'.
+   */
+  codex?: {
+    model?: string
+    baseUrl?: string
+  }
 }
 
 // ==================== GenerateRouter ====================
@@ -138,15 +146,18 @@ export class GenerateRouter {
   constructor(
     private vercel: AIProvider,
     private agentSdk: AIProvider | null = null,
+    private codex: AIProvider | null = null,
   ) {}
 
   /** Resolve the active provider, optionally overridden per-request. */
   async resolve(override?: string): Promise<AIProvider> {
+    if (override === 'codex' && this.codex) return this.codex
     // 'claude-code' is a legacy alias for 'agent-sdk'
     if ((override === 'agent-sdk' || override === 'claude-code') && this.agentSdk) return this.agentSdk
     if (override === 'vercel-ai-sdk') return this.vercel
 
     const config = await readAIProviderConfig()
+    if (config.backend === 'codex' && this.codex) return this.codex
     if ((config.backend === 'agent-sdk' || config.backend === 'claude-code') && this.agentSdk) return this.agentSdk
     return this.vercel
   }
diff --git a/src/core/config.ts b/src/core/config.ts
index 0a496dd7..32a514f9 100644
--- a/src/core/config.ts
+++ b/src/core/config.ts
@@ -16,7 +16,7 @@ const engineSchema = z.object({
 const loginMethodSchema = z.enum(['api-key', 'claudeai'])
 
 export const aiProviderSchema = z.object({
-  backend: z.enum(['claude-code', 'vercel-ai-sdk', 'agent-sdk']).default('claude-code'),
+  backend: z.enum(['claude-code', 'vercel-ai-sdk', 'agent-sdk', 'codex']).default('claude-code'),
   provider: z.string().default('anthropic'),
   model: z.string().default('claude-sonnet-4-6'),
   baseUrl: z.string().min(1).optional(),
@@ -192,6 +192,11 @@ export const agentSdkOverrideSchema = z.object({
   loginMethod: loginMethodSchema.optional(),
 })
 
+export const codexOverrideSchema = z.object({
+  model: z.string().optional(),
+  baseUrl: z.string().optional(),
+})
+
 export const webSubchannelSchema = z.object({
   /** URL-safe identifier. Used as session path segment: data/sessions/web/{id}.jsonl */
   id: z.string().regex(/^[a-z0-9-_]+$/, 'id must be lowercase alphanumeric with hyphens/underscores'),
@@ -199,11 +204,13 @@ export const webSubchannelSchema = z.object({
   /** System prompt override for this channel. */
   systemPrompt: z.string().optional(),
   /** AI backend override. Falls back to global config if omitted. */
-  provider: z.enum(['claude-code', 'vercel-ai-sdk', 'agent-sdk']).optional(),
+  provider: z.enum(['claude-code', 'vercel-ai-sdk', 'agent-sdk', 'codex']).optional(),
   /** Vercel AI SDK model override. Only used when provider is 'vercel-ai-sdk'. */
   vercelAiSdk: vercelAiSdkOverrideSchema.optional(),
   /** Agent SDK model override. Only used when provider is 'agent-sdk'. */
   agentSdk: agentSdkOverrideSchema.optional(),
+  /** Codex model override. Only used when provider is 'codex'. */
+  codex: codexOverrideSchema.optional(),
   /** Tool names to disable in addition to the global disabled list. */
   disabledTools: z.array(z.string()).optional(),
 })
@@ -445,7 +452,7 @@ export async function readConnectorsConfig() {
 
 // ==================== AI Backend Helpers ====================
 
-export type AIBackend = 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk'
+export type AIBackend = 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk' | 'codex'
 
 /** Read the current AI backend from ai-provider-manager.json. */
 export async function readAIBackend(): Promise<{ backend: AIBackend }> {
diff --git a/src/core/session.ts b/src/core/session.ts
index cb2e8499..b0a94444 100644
--- a/src/core/session.ts
+++ b/src/core/session.ts
@@ -36,7 +36,7 @@ export interface SessionEntry {
   sessionId: string
   timestamp: string
   /** Which provider generated this entry. */
-  provider?: 'vercel-ai' | 'claude-code' | 'agent-sdk' | 'human' | 'compaction' | 'notification'
+  provider?: 'vercel-ai' | 'claude-code' | 'agent-sdk' | 'codex' | 'human' | 'compaction' | 'notification'
   cwd?: string
   /** Arbitrary metadata attached to the entry (e.g. { kind: 'notification', source: 'heartbeat' }). */
   metadata?: Record<string, unknown>
diff --git a/src/main.ts b/src/main.ts
index a4d03941..9f3421ce 100644
--- a/src/main.ts
+++ b/src/main.ts
@@ -34,6 +34,7 @@ import { AgentCenter } from './core/agent-center.js'
 import { GenerateRouter } from './core/ai-provider-manager.js'
 import { VercelAIProvider } from './ai-providers/vercel-ai-sdk/vercel-provider.js'
 import { AgentSdkProvider } from './ai-providers/agent-sdk/agent-sdk-provider.js'
+import { CodexProvider } from './ai-providers/codex/index.js'
 import { createEventLog } from './core/event-log.js'
 import { createToolCallLog } from './core/tool-call-log.js'
 import { createCronEngine, createCronListener, createCronTools } from './task/cron/index.js'
@@ -223,7 +224,11 @@ async function main() {
     () => toolCenter.getVercelTools(),
     getInstructions,
   )
-  const router = new GenerateRouter(vercelProvider, agentSdkProvider)
+  const codexProvider = new CodexProvider(
+    () => toolCenter.getVercelTools(),
+    getInstructions,
+  )
+  const router = new GenerateRouter(vercelProvider, agentSdkProvider, codexProvider)
 
   const agentCenter = new AgentCenter({
     router,

From 4d2d5ceb8d41ed2c2a73896253cc1466f4e7e383 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 19:35:44 +0800
Subject: [PATCH 08/15] feat: Codex provider UI + dual auth mode (OAuth / API
 key)

Frontend:
- Replace legacy "OpenAI" (Vercel wrapper) card with native "OpenAI / Codex"
- Add CodexAuthForm with ChatGPT Subscription / API Key toggle
- Add Codex config section in ChannelConfigModal (model, baseUrl, apiKey, loginMethod)
- Add CodexOverride type + update provider unions across frontend

Backend:
- Codex provider supports dual auth: codex-oauth (default) reads ~/.codex/auth.json,
  api-key mode uses apiKeys.openai from config
- Add loginMethod to codexOverrideSchema and GenerateOpts
- Fix /ai-provider PUT validation to accept 'codex' backend

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 src/ai-providers/codex/codex-provider.ts |  32 ++++--
 src/ai-providers/types.ts                |   2 +-
 src/connectors/web/routes/config.ts      |   4 +-
 src/core/ai-provider-manager.ts          |   2 +
 src/core/config.ts                       |   4 +-
 ui/src/api/channels.ts                   |   5 +-
 ui/src/api/types.ts                      |  12 ++-
 ui/src/components/ChannelConfigModal.tsx |  73 +++++++++++++-
 ui/src/pages/AIProviderPage.tsx          | 119 ++++++++++++++++++++---
 9 files changed, 218 insertions(+), 35 deletions(-)

diff --git a/src/ai-providers/codex/codex-provider.ts b/src/ai-providers/codex/codex-provider.ts
index 0125aad2..49d27bff 100644
--- a/src/ai-providers/codex/codex-provider.ts
+++ b/src/ai-providers/codex/codex-provider.ts
@@ -25,8 +25,9 @@ const logger = pino({
   transport: { target: 'pino/file', options: { destination: 'logs/codex.log', mkdir: true } },
 })
 
-const DEFAULT_BASE_URL = 'https://chatgpt.com/backend-api/codex'
-const DEFAULT_MODEL = 'codex-mini-latest'
+const DEFAULT_OAUTH_BASE_URL = 'https://chatgpt.com/backend-api/codex'
+const DEFAULT_API_BASE_URL = 'https://api.openai.com/v1'
+const DEFAULT_MODEL = 'gpt-5.4'
 
 // ==================== Provider ====================
 
@@ -38,19 +39,30 @@ export class CodexProvider implements AIProvider {
     private getSystemPrompt: () => Promise<string>,
   ) {}
 
-  /** Create an OpenAI client with the current access token. */
+  /**
+   * Create an OpenAI client.
+   *
+   * - loginMethod 'codex-oauth' (default): reads ~/.codex/auth.json, hits
+   *   ChatGPT subscription endpoint. Usage billed to ChatGPT plan.
+   * - loginMethod 'api-key': uses apiKeys.openai from config (or override).
+   *   Standard OpenAI API billing, or compatible third-party endpoint.
+   */
   private async createClient(opts?: GenerateOpts): Promise<{ client: OpenAI; model: string }> {
-    const token = await getAccessToken()
     const aiConfig = await readAIProviderConfig()
-    const baseURL = opts?.codex?.baseUrl ?? DEFAULT_BASE_URL
     const model = opts?.codex?.model ?? aiConfig.model ?? DEFAULT_MODEL
+    const loginMethod = opts?.codex?.loginMethod ?? aiConfig.loginMethod ?? 'codex-oauth'
 
-    const client = new OpenAI({
-      apiKey: token,
-      baseURL,
-    })
+    if (loginMethod === 'api-key') {
+      const apiKey = opts?.codex?.apiKey ?? aiConfig.apiKeys.openai
+      if (!apiKey) throw new Error('Codex api-key mode requires apiKeys.openai in config.')
+      const baseURL = opts?.codex?.baseUrl ?? aiConfig.baseUrl ?? DEFAULT_API_BASE_URL
+      return { client: new OpenAI({ apiKey, baseURL }), model }
+    }
 
-    return { client, model }
+    // OAuth mode
+    const token = await getAccessToken()
+    const baseURL = opts?.codex?.baseUrl ?? DEFAULT_OAUTH_BASE_URL
+    return { client: new OpenAI({ apiKey: token, baseURL }), model }
   }
 
   async ask(prompt: string): Promise<ProviderResult> {
diff --git a/src/ai-providers/types.ts b/src/ai-providers/types.ts
index b31ece4a..59d75bb7 100644
--- a/src/ai-providers/types.ts
+++ b/src/ai-providers/types.ts
@@ -32,7 +32,7 @@ export interface GenerateOpts {
   disabledTools?: string[]
   vercelAiSdk?: { provider: string; model: string; baseUrl?: string; apiKey?: string }
   agentSdk?: { model?: string; apiKey?: string; baseUrl?: string }
-  codex?: { model?: string; baseUrl?: string }
+  codex?: { model?: string; baseUrl?: string; apiKey?: string; loginMethod?: 'api-key' | 'codex-oauth' }
 }
 
 // ==================== AIProvider ====================
diff --git a/src/connectors/web/routes/config.ts b/src/connectors/web/routes/config.ts
index 0a9738fd..26036109 100644
--- a/src/connectors/web/routes/config.ts
+++ b/src/connectors/web/routes/config.ts
@@ -23,8 +23,8 @@ export function createConfigRoutes(opts?: ConfigRouteOpts) {
     try {
       const body = await c.req.json<{ backend?: string }>()
       const backend = body.backend
-      if (backend !== 'claude-code' && backend !== 'vercel-ai-sdk' && backend !== 'agent-sdk') {
-        return c.json({ error: 'Invalid backend. Must be "claude-code", "vercel-ai-sdk", or "agent-sdk".' }, 400)
+      if (backend !== 'claude-code' && backend !== 'vercel-ai-sdk' && backend !== 'agent-sdk' && backend !== 'codex') {
+        return c.json({ error: 'Invalid backend. Must be "claude-code", "vercel-ai-sdk", "agent-sdk", or "codex".' }, 400)
       }
       await writeAIBackend(backend as AIBackend)
       return c.json({ backend })
diff --git a/src/core/ai-provider-manager.ts b/src/core/ai-provider-manager.ts
index 65a615e0..ee3342f3 100644
--- a/src/core/ai-provider-manager.ts
+++ b/src/core/ai-provider-manager.ts
@@ -136,6 +136,8 @@ export interface AskOptions {
   codex?: {
     model?: string
     baseUrl?: string
+    apiKey?: string
+    loginMethod?: 'api-key' | 'codex-oauth'
   }
 }
 
diff --git a/src/core/config.ts b/src/core/config.ts
index 32a514f9..46508c7d 100644
--- a/src/core/config.ts
+++ b/src/core/config.ts
@@ -13,7 +13,7 @@ const engineSchema = z.object({
   port: z.number().int().positive().default(3000),
 })
 
-const loginMethodSchema = z.enum(['api-key', 'claudeai'])
+const loginMethodSchema = z.enum(['api-key', 'claudeai', 'codex-oauth'])
 
 export const aiProviderSchema = z.object({
   backend: z.enum(['claude-code', 'vercel-ai-sdk', 'agent-sdk', 'codex']).default('claude-code'),
@@ -195,6 +195,8 @@ export const agentSdkOverrideSchema = z.object({
 export const codexOverrideSchema = z.object({
   model: z.string().optional(),
   baseUrl: z.string().optional(),
+  apiKey: z.string().optional(),
+  loginMethod: z.enum(['api-key', 'codex-oauth']).optional(),
 })
 
 export const webSubchannelSchema = z.object({
diff --git a/ui/src/api/channels.ts b/ui/src/api/channels.ts
index 1627a69a..3479ad4c 100644
--- a/ui/src/api/channels.ts
+++ b/ui/src/api/channels.ts
@@ -1,13 +1,14 @@
 import { headers } from './client'
-import type { WebChannel, VercelAiSdkOverride, AgentSdkOverride } from './types'
+import type { WebChannel, VercelAiSdkOverride, AgentSdkOverride, CodexOverride } from './types'
 
 export interface ChannelListItem {
   id: string
   label: string
   systemPrompt?: string
-  provider?: 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk'
+  provider?: 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk' | 'codex'
   vercelAiSdk?: VercelAiSdkOverride
   agentSdk?: AgentSdkOverride
+  codex?: CodexOverride
   disabledTools?: string[]
 }
 
diff --git a/ui/src/api/types.ts b/ui/src/api/types.ts
index 7d9ab8db..03f9c946 100644
--- a/ui/src/api/types.ts
+++ b/ui/src/api/types.ts
@@ -7,7 +7,7 @@ export interface VercelAiSdkOverride {
   apiKey?: string
 }
 
-export type LoginMethod = 'api-key' | 'claudeai'
+export type LoginMethod = 'api-key' | 'claudeai' | 'codex-oauth'
 
 export interface AgentSdkOverride {
   model?: string
@@ -16,13 +16,21 @@ export interface AgentSdkOverride {
   loginMethod?: LoginMethod
 }
 
+export interface CodexOverride {
+  model?: string
+  baseUrl?: string
+  apiKey?: string
+  loginMethod?: 'api-key' | 'codex-oauth'
+}
+
 export interface WebChannel {
   id: string
   label: string
   systemPrompt?: string
-  provider?: 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk'
+  provider?: 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk' | 'codex'
   vercelAiSdk?: VercelAiSdkOverride
   agentSdk?: AgentSdkOverride
+  codex?: CodexOverride
   disabledTools?: string[]
 }
 
diff --git a/ui/src/components/ChannelConfigModal.tsx b/ui/src/components/ChannelConfigModal.tsx
index 46de7728..73b8a0a0 100644
--- a/ui/src/components/ChannelConfigModal.tsx
+++ b/ui/src/components/ChannelConfigModal.tsx
@@ -31,8 +31,15 @@ export function ChannelConfigModal({ channel, onClose, onSaved }: ChannelConfigM
   const [aApiKey, setAApiKey] = useState(channel.agentSdk?.apiKey ?? '')
   const [aLoginMethod, setALoginMethod] = useState<LoginMethod | ''>(channel.agentSdk?.loginMethod ?? '')
 
+  // Codex override state
+  const [cModel, setCModel] = useState(channel.codex?.model ?? '')
+  const [cBaseUrl, setCBaseUrl] = useState(channel.codex?.baseUrl ?? '')
+  const [cApiKey, setCApiKey] = useState(channel.codex?.apiKey ?? '')
+  const [cLoginMethod, setCLoginMethod] = useState<'api-key' | 'codex-oauth' | ''>(channel.codex?.loginMethod ?? '')
+
   const showVercelConfig = provider === 'vercel-ai-sdk'
   const showAgentSdkConfig = provider === 'agent-sdk'
+  const showCodexConfig = provider === 'codex'
 
   useEffect(() => {
     api.tools.load().then(({ inventory }) => setTools(inventory)).catch(() => {})
@@ -60,12 +67,22 @@ export function ChannelConfigModal({ channel, onClose, onSaved }: ChannelConfigM
           }
         : undefined
 
+      const codex = showCodexConfig && (cModel || cLoginMethod)
+        ? {
+            ...(cModel ? { model: cModel } : {}),
+            ...(cBaseUrl ? { baseUrl: cBaseUrl } : {}),
+            ...(cApiKey ? { apiKey: cApiKey } : {}),
+            ...(cLoginMethod ? { loginMethod: cLoginMethod as 'api-key' | 'codex-oauth' } : {}),
+          }
+        : undefined
+
       const { channel: updated } = await api.channels.update(channel.id, {
         label: label.trim() || channel.label,
         systemPrompt: systemPrompt.trim() || undefined,
-        provider: (provider as 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk') || undefined,
+        provider: (provider as 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk' | 'codex') || undefined,
         vercelAiSdk: vercelAiSdk ?? (null as unknown as undefined),
         agentSdk: agentSdk ?? (null as unknown as undefined),
+        codex: codex ?? (null as unknown as undefined),
         disabledTools: disabledTools.size > 0 ? [...disabledTools] : undefined,
       })
       onSaved(updated)
@@ -151,6 +168,7 @@ export function ChannelConfigModal({ channel, onClose, onSaved }: ChannelConfigM
               <option value="">Default (global)</option>
               <option value="agent-sdk">Claude</option>
               <option value="vercel-ai-sdk">Vercel AI SDK</option>
+              <option value="codex">Codex</option>
             </select>
           </div>
 
@@ -267,6 +285,59 @@ export function ChannelConfigModal({ channel, onClose, onSaved }: ChannelConfigM
             </div>
           )}
 
+          {/* Codex config — only when provider is codex */}
+          {showCodexConfig && (
+            <div className="rounded-lg border border-border/50 bg-bg-secondary/30 p-3 space-y-3">
+              <p className="text-xs font-medium text-text-muted">Codex Override</p>
+
+              <div>
+                <label className="block text-xs text-text-muted/70 mb-1">Login Method</label>
+                <select
+                  value={cLoginMethod}
+                  onChange={(e) => setCLoginMethod((e.target.value || '') as 'api-key' | 'codex-oauth' | '')}
+                  className={inputClass}
+                >
+                  <option value="">Default (global)</option>
+                  <option value="codex-oauth">ChatGPT Subscription</option>
+                  <option value="api-key">API Key</option>
+                </select>
+              </div>
+
+              <div>
+                <label className="block text-xs text-text-muted/70 mb-1">Model</label>
+                <input
+                  type="text"
+                  value={cModel}
+                  onChange={(e) => setCModel(e.target.value)}
+                  placeholder="e.g. gpt-5.4"
+                  className={inputClass}
+                />
+              </div>
+
+              <div>
+                <label className="block text-xs text-text-muted/70 mb-1">Base URL <span className="text-text-muted/40">(optional)</span></label>
+                <input
+                  type="text"
+                  value={cBaseUrl}
+                  onChange={(e) => setCBaseUrl(e.target.value)}
+                  placeholder="Leave empty for default"
+                  className={inputClass}
+                />
+              </div>
+
+              <div>
+                <label className="block text-xs text-text-muted/70 mb-1">API Key <span className="text-text-muted/40">(optional, overrides global)</span></label>
+                <input
+                  type="password"
+                  value={cApiKey}
+                  onChange={(e) => setCApiKey(e.target.value)}
+                  placeholder="sk-..."
+                  className={inputClass}
+                />
+              </div>
+            </div>
+          )}
+
           {/* Disabled Tools */}
           <div>
             <label className="block text-xs font-medium text-text-muted mb-2">
diff --git a/ui/src/pages/AIProviderPage.tsx b/ui/src/pages/AIProviderPage.tsx
index 13a474eb..6728466e 100644
--- a/ui/src/pages/AIProviderPage.tsx
+++ b/ui/src/pages/AIProviderPage.tsx
@@ -49,19 +49,21 @@ function detectCustomMode(provider: string, model: string): boolean {
   return !presets.some((p) => p.value === model)
 }
 
-/** UI-level backend. 'openai' is a facade over vercel-ai-sdk with provider=openai. */
-type UIBackend = 'agent-sdk' | 'openai' | 'vercel-ai-sdk'
+/** UI-level backend selector. */
+type UIBackend = 'agent-sdk' | 'codex' | 'vercel-ai-sdk'
 
 /** Default provider/model per UI backend — applied on every switch to avoid stale config. */
 const BACKEND_DEFAULTS: Record<UIBackend, { backend: string; provider: string; model: string }> = {
   'agent-sdk':    { backend: 'agent-sdk',    provider: 'anthropic', model: 'claude-sonnet-4-6' },
-  'openai':       { backend: 'vercel-ai-sdk', provider: 'openai',   model: PROVIDER_MODELS.openai[0].value },
+  'codex':        { backend: 'codex',        provider: 'openai',   model: 'gpt-5.4' },
   'vercel-ai-sdk': { backend: 'vercel-ai-sdk', provider: 'anthropic', model: PROVIDER_MODELS.anthropic[0].value },
 }
 
 /** Derive initial UI backend from config. */
 function detectUIBackend(config: AIProviderConfig): UIBackend {
-  if (config.backend === 'vercel-ai-sdk' && config.provider === 'openai') return 'openai'
+  if (config.backend === 'codex') return 'codex'
+  // Legacy: openai via vercel-ai-sdk → migrate to codex
+  if (config.backend === 'vercel-ai-sdk' && config.provider === 'openai') return 'codex'
   if (config.backend === 'vercel-ai-sdk') return 'vercel-ai-sdk'
   return 'agent-sdk'
 }
@@ -133,11 +135,11 @@ export function AIProviderPage() {
                   description="Claude Code login or Anthropic API key"
                 />
                 <BackendCard
-                  selected={uiBackend === 'openai'}
-                  onClick={() => handleBackendSwitch('openai')}
-                  icon={<svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><circle cx="12" cy="12" r="3" /><path d="M12 2v4M12 18v4M4.93 4.93l2.83 2.83M16.24 16.24l2.83 2.83M2 12h4M18 12h4M4.93 19.07l2.83-2.83M16.24 7.76l2.83-2.83" /></svg>}
-                  title="OpenAI"
-                  description="GPT models via OpenAI API"
+                  selected={uiBackend === 'codex'}
+                  onClick={() => handleBackendSwitch('codex')}
+                  icon={<svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><polyline points="16 18 22 12 16 6" /><polyline points="8 6 2 12 8 18" /><line x1="14" y1="4" x2="10" y2="20" /></svg>}
+                  title="OpenAI / Codex"
+                  description="ChatGPT subscription or OpenAI API key"
                 />
                 <BackendCard
                   selected={uiBackend === 'vercel-ai-sdk'}
@@ -156,19 +158,19 @@ export function AIProviderPage() {
               </ConfigSection>
             )}
 
-            {/* OpenAI simplified form */}
-            {uiBackend === 'openai' && (
-              <ConfigSection title="Model" description="Select a model and enter your OpenAI API key.">
-                <OpenAIForm aiProvider={config.aiProvider} />
-              </ConfigSection>
-            )}
-
             {/* Full model form (only for Vercel AI SDK) */}
             {uiBackend === 'vercel-ai-sdk' && (
               <ConfigSection title="Model" description="Provider, model, and API keys. Changes take effect on the next request.">
                 <ModelForm aiProvider={config.aiProvider} />
               </ConfigSection>
             )}
+
+            {/* Codex auth + model form */}
+            {uiBackend === 'codex' && (
+              <ConfigSection title="Authentication" description="Choose how Alice connects to OpenAI / Codex models.">
+                <CodexAuthForm aiProvider={config.aiProvider} onUpdate={(patch) => setConfig((c) => c ? { ...c, aiProvider: { ...c.aiProvider, ...patch } } : c)} />
+              </ConfigSection>
+            )}
           </div>
       </div>
       ) : (
@@ -561,6 +563,91 @@ function ModelForm({ aiProvider }: { aiProvider: AIProviderConfig }) {
 
 // ==================== Agent SDK Auth Form ====================
 
+const CODEX_LOGIN_METHODS: { value: LoginMethod; label: string; subtitle: string; hint: string }[] = [
+  { value: 'codex-oauth', label: 'ChatGPT Subscription', subtitle: 'Use your ChatGPT plan', hint: 'Requires local Codex CLI login (run codex login in terminal). Usage billed to your ChatGPT subscription.' },
+  { value: 'api-key', label: 'API Key', subtitle: 'Pay per token', hint: 'Enter your OpenAI API key below. Billed per token to your API account.' },
+]
+
+function CodexAuthForm({ aiProvider, onUpdate }: { aiProvider: AIProviderConfig; onUpdate: (patch: Partial<AIProviderConfig>) => void }) {
+  const [loginMethod, setLoginMethod] = useState<LoginMethod>(aiProvider.loginMethod ?? 'codex-oauth')
+  const [apiKey, setApiKey] = useState('')
+  const [keySaveStatus, setKeySaveStatus] = useState<SaveStatus>('idle')
+  const savedTimer = useRef<ReturnType<typeof setTimeout> | null>(null)
+
+  useEffect(() => () => { if (savedTimer.current) clearTimeout(savedTimer.current) }, [])
+
+  const handleLoginMethodChange = async (method: LoginMethod) => {
+    setLoginMethod(method)
+    try {
+      await api.config.updateSection('aiProvider', { ...aiProvider, loginMethod: method })
+      onUpdate({ loginMethod: method })
+    } catch { setLoginMethod(loginMethod) }
+  }
+
+  const handleSaveKey = async () => {
+    if (!apiKey) return
+    setKeySaveStatus('saving')
+    try {
+      const updatedKeys = { ...aiProvider.apiKeys, openai: apiKey }
+      await api.config.updateSection('aiProvider', { ...aiProvider, apiKeys: updatedKeys })
+      onUpdate({ apiKeys: updatedKeys })
+      setApiKey('')
+      setKeySaveStatus('saved')
+      if (savedTimer.current) clearTimeout(savedTimer.current)
+      savedTimer.current = setTimeout(() => setKeySaveStatus('idle'), 2000)
+    } catch { setKeySaveStatus('error') }
+  }
+
+  return (
+    <>
+      <div className="grid grid-cols-2 gap-3">
+        {CODEX_LOGIN_METHODS.map((m) => (
+          <BackendCard
+            key={m.value}
+            selected={loginMethod === m.value}
+            onClick={() => handleLoginMethodChange(m.value)}
+            icon={m.value === 'codex-oauth'
+              ? <svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><path d="M15 3h4a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2h-4" /><polyline points="10 17 15 12 10 7" /><line x1="15" y1="12" x2="3" y2="12" /></svg>
+              : <svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><rect x="3" y="11" width="18" height="11" rx="2" ry="2" /><path d="M7 11V7a5 5 0 0 1 10 0v4" /></svg>}
+            title={m.label}
+            description={m.subtitle}
+          />
+        ))}
+      </div>
+      <p className="text-[11px] text-text-muted mt-2">
+        {CODEX_LOGIN_METHODS.find((m) => m.value === loginMethod)?.hint}
+      </p>
+
+      {loginMethod === 'api-key' && (
+        <Field label="OpenAI API Key">
+          <div className="relative">
+            <input
+              className={inputClass}
+              type="password"
+              value={apiKey}
+              onChange={(e) => setApiKey(e.target.value)}
+              placeholder={aiProvider.apiKeys?.openai ? '(configured)' : 'sk-...'}
+            />
+            {aiProvider.apiKeys?.openai && (
+              <span className="absolute right-2.5 top-1/2 -translate-y-1/2 text-[10px] text-green">active</span>
+            )}
+          </div>
+          <div className="flex items-center gap-3 mt-2">
+            <button
+              onClick={handleSaveKey}
+              disabled={!apiKey || keySaveStatus === 'saving'}
+              className="btn-primary"
+            >
+              Save Key
+            </button>
+            <SaveIndicator status={keySaveStatus} onRetry={handleSaveKey} />
+          </div>
+        </Field>
+      )}
+    </>
+  )
+}
+
 function AgentSdkAuthForm({ aiProvider, onUpdate }: { aiProvider: AIProviderConfig; onUpdate: (patch: Partial<AIProviderConfig>) => void }) {
   const [loginMethod, setLoginMethod] = useState<LoginMethod>(aiProvider.loginMethod ?? 'api-key')
   const [apiKey, setApiKey] = useState('')

From 37c2336232a4205a5d34e1fa2498575ef371df75 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 19:39:41 +0800
Subject: [PATCH 09/15] fix: add codex provider support to channels CRUD
 endpoints

POST and PUT /channels were rejecting 'codex' as provider and not
passing the codex override object through to config.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 src/connectors/web/routes/channels.ts | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/src/connectors/web/routes/channels.ts b/src/connectors/web/routes/channels.ts
index 120aaf01..7d1b40bc 100644
--- a/src/connectors/web/routes/channels.ts
+++ b/src/connectors/web/routes/channels.ts
@@ -32,6 +32,7 @@ export function createChannelsRoutes({ sessions, sseByChannel }: ChannelsDeps) {
       provider?: string
       vercelAiSdk?: { provider: string; model: string; baseUrl?: string; apiKey?: string }
       agentSdk?: { model?: string; apiKey?: string; baseUrl?: string }
+      codex?: { model?: string; baseUrl?: string; apiKey?: string; loginMethod?: 'api-key' | 'codex-oauth' }
       disabledTools?: string[]
     }
 
@@ -54,13 +55,14 @@ export function createChannelsRoutes({ sessions, sseByChannel }: ChannelsDeps) {
       id: body.id,
       label: body.label.trim(),
       ...(body.systemPrompt ? { systemPrompt: body.systemPrompt } : {}),
-      ...(body.provider === 'claude-code' || body.provider === 'vercel-ai-sdk' || body.provider === 'agent-sdk'
+      ...(body.provider === 'claude-code' || body.provider === 'vercel-ai-sdk' || body.provider === 'agent-sdk' || body.provider === 'codex'
         ? { provider: body.provider }
         : {}),
       ...(body.vercelAiSdk?.provider && body.vercelAiSdk?.model
         ? { vercelAiSdk: body.vercelAiSdk }
         : {}),
       ...(body.agentSdk ? { agentSdk: body.agentSdk } : {}),
+      ...(body.codex ? { codex: body.codex } : {}),
       ...(body.disabledTools?.length ? { disabledTools: body.disabledTools } : {}),
     }
 
@@ -86,6 +88,7 @@ export function createChannelsRoutes({ sessions, sseByChannel }: ChannelsDeps) {
       provider?: string
       vercelAiSdk?: { provider: string; model: string; baseUrl?: string; apiKey?: string } | null
       agentSdk?: { model?: string; apiKey?: string; baseUrl?: string } | null
+      codex?: { model?: string; baseUrl?: string; apiKey?: string; loginMethod?: 'api-key' | 'codex-oauth' } | null
       disabledTools?: string[]
     }
 
@@ -97,7 +100,7 @@ export function createChannelsRoutes({ sessions, sseByChannel }: ChannelsDeps) {
       ...existing[idx],
       ...(body.label !== undefined ? { label: body.label } : {}),
       ...(body.systemPrompt !== undefined ? { systemPrompt: body.systemPrompt || undefined } : {}),
-      ...(body.provider === 'claude-code' || body.provider === 'vercel-ai-sdk' || body.provider === 'agent-sdk'
+      ...(body.provider === 'claude-code' || body.provider === 'vercel-ai-sdk' || body.provider === 'agent-sdk' || body.provider === 'codex'
         ? { provider: body.provider }
         : body.provider === null || body.provider === ''
           ? { provider: undefined }
@@ -108,6 +111,9 @@ export function createChannelsRoutes({ sessions, sseByChannel }: ChannelsDeps) {
       ...(body.agentSdk !== undefined
         ? { agentSdk: body.agentSdk ?? undefined }
         : {}),
+      ...(body.codex !== undefined
+        ? { codex: body.codex ?? undefined }
+        : {}),
       ...(body.disabledTools !== undefined ? { disabledTools: body.disabledTools?.length ? body.disabledTools : undefined } : {}),
     }
     existing[idx] = updated

From 5e4b24c9a4455785b55334170b62e008cb1f4a00 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 19:55:02 +0800
Subject: [PATCH 10/15] fix: improve codex provider error logging with request
 details

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 src/ai-providers/codex/codex-provider.ts | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/src/ai-providers/codex/codex-provider.ts b/src/ai-providers/codex/codex-provider.ts
index 49d27bff..589353e9 100644
--- a/src/ai-providers/codex/codex-provider.ts
+++ b/src/ai-providers/codex/codex-provider.ts
@@ -168,7 +168,15 @@ export class CodexProvider implements AIProvider {
           yield { type: 'done', result: { text: errorText, media: [] } }
           return
         }
-        logger.error({ err: err?.message, status: err?.status }, 'responses_api_error')
+        logger.error({
+          err: err?.message,
+          status: err?.status,
+          headers: err?.headers,
+          body: err?.error ?? err?.body,
+          model,
+          inputItems: input.length,
+          toolCount: tools.length,
+        }, 'responses_api_error')
         const errorText = accumulatedText + stepText +
           `\n\n[Codex API error: ${err?.message ?? 'unknown error'}]`
         yield { type: 'done', result: { text: errorText, media: [] } }

From 5a837c189f7cc2a24bee861a6d9df521aec3cf01 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 20:28:19 +0800
Subject: [PATCH 11/15] refactor: replace flat AI config with profile-based
 system
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The old ai-provider-manager.json was a flat object where model/loginMethod/baseUrl
were shared across all backends — switching from Claude to Codex would send
"claude-sonnet-4-6" to OpenAI (400 error).

New structure separates selection from configuration:

  { apiKeys: {...}, profiles: { "slug": { backend, label, model, ... } }, activeProfile: "slug" }

- Each profile carries only the fields its backend needs (discriminated union)
- Profiles are identified by slug, referenced by main channel and subchannels
- Global apiKeys shared, per-profile apiKey override supported
- Auto-migration from old flat format on first load
- Subchannels simplified: `profile: "slug"` replaces inline provider/vercelAiSdk/agentSdk/codex
- Providers receive ResolvedProfile instead of reading flat config directly
- GenerateRouter resolves profile → picks provider by backend
- Profile CRUD API: GET/POST/PUT/DELETE /config/profiles, PUT /config/active-profile
- Telegram settings menu shows profiles instead of backends

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 .../agent-sdk/agent-sdk-provider.ts           |  15 +-
 src/ai-providers/agent-sdk/query.ts           |  13 +-
 src/ai-providers/codex/codex-provider.ts      |  24 +-
 src/ai-providers/types.ts                     |   6 +-
 .../vercel-ai-sdk/model-factory.ts            |  39 +--
 .../vercel-ai-sdk/vercel-provider.spec.ts     |  23 +-
 .../vercel-ai-sdk/vercel-provider.ts          |  15 +-
 src/connectors/telegram/telegram-plugin.ts    |  59 ++--
 .../web/__tests__/chat-streaming.spec.ts      |   5 +
 src/connectors/web/routes/channels.ts         |  34 +--
 src/connectors/web/routes/chat.ts             |   5 +-
 src/connectors/web/routes/config.ts           | 117 ++++++--
 src/core/__tests__/pipeline/helpers.ts        |   7 +
 src/core/__tests__/pipeline/streaming.spec.ts |   5 +
 src/core/agent-center.spec.ts                 |  11 +-
 src/core/agent-center.ts                      |   8 +-
 src/core/ai-provider-manager.spec.ts          |  56 ++--
 src/core/ai-provider-manager.ts               | 102 ++-----
 src/core/config.spec.ts                       | 133 +++++----
 src/core/config.ts                            | 270 ++++++++++++++----
 20 files changed, 560 insertions(+), 387 deletions(-)

diff --git a/src/ai-providers/agent-sdk/agent-sdk-provider.ts b/src/ai-providers/agent-sdk/agent-sdk-provider.ts
index f5991ca1..723e9369 100644
--- a/src/ai-providers/agent-sdk/agent-sdk-provider.ts
+++ b/src/ai-providers/agent-sdk/agent-sdk-provider.ts
@@ -15,7 +15,7 @@ import type { SessionEntry } from '../../core/session.js'
 import type { AgentSdkConfig, AgentSdkOverride } from './query.js'
 import { toTextHistory } from '../../core/session.js'
 import { buildChatHistoryPrompt, DEFAULT_MAX_HISTORY } from '../utils.js'
-import { readAgentConfig } from '../../core/config.js'
+import { readAgentConfig, resolveProfile } from '../../core/config.js'
 import { createChannel } from '../../core/async-channel.js'
 import { askAgentSdk } from './query.js'
 import { buildAgentSdkMcpServer } from './tool-bridge.js'
@@ -47,8 +47,13 @@ export class AgentSdkProvider implements AIProvider {
   async ask(prompt: string): Promise<ProviderResult> {
     const config = await this.resolveConfig()
     config.systemPrompt = await this.getSystemPrompt()
+    const profile = await resolveProfile()
+    const override: AgentSdkOverride = {
+      model: profile.model, apiKey: profile.apiKey, baseUrl: profile.baseUrl,
+      loginMethod: profile.loginMethod as 'api-key' | 'claudeai' | undefined,
+    }
     const mcpServer = await this.buildMcpServer()
-    const result = await askAgentSdk(prompt, config, undefined, mcpServer)
+    const result = await askAgentSdk(prompt, config, override, mcpServer)
     return { text: result.text, media: [] }
   }
 
@@ -66,7 +71,11 @@ export class AgentSdkProvider implements AIProvider {
       systemPrompt: opts?.systemPrompt ?? await this.getSystemPrompt(),
     }
 
-    const override: AgentSdkOverride | undefined = opts?.agentSdk
+    // Build override from resolved profile
+    const profile = opts?.profile
+    const override: AgentSdkOverride | undefined = profile
+      ? { model: profile.model, apiKey: profile.apiKey, baseUrl: profile.baseUrl, loginMethod: profile.loginMethod as 'api-key' | 'claudeai' | undefined }
+      : undefined
     const mcpServer = await this.buildMcpServer(opts?.disabledTools)
 
     const channel = createChannel<ProviderEvent>()
diff --git a/src/ai-providers/agent-sdk/query.ts b/src/ai-providers/agent-sdk/query.ts
index 6136e33c..d4ce05d1 100644
--- a/src/ai-providers/agent-sdk/query.ts
+++ b/src/ai-providers/agent-sdk/query.ts
@@ -10,7 +10,7 @@ import type { McpSdkServerConfigWithInstance } from '@anthropic-ai/claude-agent-
 import { pino } from 'pino'
 import type { ContentBlock } from '../../core/session.js'
 
-import { readAIProviderConfig } from '../../core/config.js'
+// Config is now resolved via profile system — override carries all needed values
 
 const logger = pino({
   transport: { target: 'pino/file', options: { destination: 'logs/agent-sdk.log', mkdir: true } },
@@ -117,9 +117,8 @@ export async function askAgentSdk(
   const finalAllowed = allowedTools.length > 0 ? allowedTools : modeAllowed
   const finalDisallowed = [...disallowedTools, ...modeDisallowed]
 
-  // Build env with authentication
-  const aiConfig = await readAIProviderConfig()
-  const loginMethod = override?.loginMethod ?? aiConfig.loginMethod ?? 'api-key'
+  // Build env with authentication — override carries resolved profile values
+  const loginMethod = override?.loginMethod ?? 'api-key'
   const isOAuthMode = loginMethod === 'claudeai'
 
   const env: Record<string, string | undefined> = { ...process.env }
@@ -127,10 +126,10 @@ export async function askAgentSdk(
     // Force OAuth by removing any inherited API key
     delete env.ANTHROPIC_API_KEY
   } else {
-    const apiKey = override?.apiKey ?? aiConfig.apiKeys.anthropic
+    const apiKey = override?.apiKey
     if (apiKey) env.ANTHROPIC_API_KEY = apiKey
   }
-  const baseUrl = override?.baseUrl ?? aiConfig.baseUrl
+  const baseUrl = override?.baseUrl
   if (baseUrl) env.ANTHROPIC_BASE_URL = baseUrl
 
   // MCP servers
@@ -149,7 +148,7 @@ export async function askAgentSdk(
       options: {
         cwd,
         env,
-        model: override?.model ?? aiConfig.model,
+        model: override?.model ?? 'claude-sonnet-4-6',
         maxTurns,
         allowedTools: finalAllowed,
         disallowedTools: finalDisallowed,
diff --git a/src/ai-providers/codex/codex-provider.ts b/src/ai-providers/codex/codex-provider.ts
index 589353e9..e13af0da 100644
--- a/src/ai-providers/codex/codex-provider.ts
+++ b/src/ai-providers/codex/codex-provider.ts
@@ -15,9 +15,10 @@ import { pino } from 'pino'
 
 import type { ProviderResult, ProviderEvent, AIProvider, GenerateOpts } from '../types.js'
 import type { SessionEntry } from '../../core/session.js'
+import type { ResolvedProfile } from '../../core/config.js'
 import { toTextHistory } from '../../core/session.js'
 import { buildChatHistoryPrompt, DEFAULT_MAX_HISTORY } from '../utils.js'
-import { readAIProviderConfig, readAgentConfig } from '../../core/config.js'
+import { readAgentConfig } from '../../core/config.js'
 import { getAccessToken, clearTokenCache } from './auth.js'
 import { convertTools } from './tool-bridge.js'
 
@@ -40,28 +41,27 @@ export class CodexProvider implements AIProvider {
   ) {}
 
   /**
-   * Create an OpenAI client.
+   * Create an OpenAI client from a resolved profile.
    *
    * - loginMethod 'codex-oauth' (default): reads ~/.codex/auth.json, hits
    *   ChatGPT subscription endpoint. Usage billed to ChatGPT plan.
-   * - loginMethod 'api-key': uses apiKeys.openai from config (or override).
+   * - loginMethod 'api-key': uses profile apiKey.
    *   Standard OpenAI API billing, or compatible third-party endpoint.
    */
-  private async createClient(opts?: GenerateOpts): Promise<{ client: OpenAI; model: string }> {
-    const aiConfig = await readAIProviderConfig()
-    const model = opts?.codex?.model ?? aiConfig.model ?? DEFAULT_MODEL
-    const loginMethod = opts?.codex?.loginMethod ?? aiConfig.loginMethod ?? 'codex-oauth'
+  private async createClient(profile?: ResolvedProfile): Promise<{ client: OpenAI; model: string }> {
+    const model = profile?.model ?? DEFAULT_MODEL
+    const loginMethod = profile?.loginMethod ?? 'codex-oauth'
 
     if (loginMethod === 'api-key') {
-      const apiKey = opts?.codex?.apiKey ?? aiConfig.apiKeys.openai
-      if (!apiKey) throw new Error('Codex api-key mode requires apiKeys.openai in config.')
-      const baseURL = opts?.codex?.baseUrl ?? aiConfig.baseUrl ?? DEFAULT_API_BASE_URL
+      const apiKey = profile?.apiKey
+      if (!apiKey) throw new Error('Codex api-key mode requires an API key. Configure it in your profile.')
+      const baseURL = profile?.baseUrl ?? DEFAULT_API_BASE_URL
       return { client: new OpenAI({ apiKey, baseURL }), model }
     }
 
     // OAuth mode
     const token = await getAccessToken()
-    const baseURL = opts?.codex?.baseUrl ?? DEFAULT_OAUTH_BASE_URL
+    const baseURL = profile?.baseUrl ?? DEFAULT_OAUTH_BASE_URL
     return { client: new OpenAI({ apiKey: token, baseURL }), model }
   }
 
@@ -99,7 +99,7 @@ export class CodexProvider implements AIProvider {
     const textHistory = toTextHistory(entries).slice(-maxHistory)
     const fullPrompt = buildChatHistoryPrompt(prompt, textHistory, opts?.historyPreamble)
 
-    const { client, model } = await this.createClient(opts)
+    const { client, model } = await this.createClient(opts?.profile)
     const instructions = opts?.systemPrompt ?? await this.getSystemPrompt()
     const agentConfig = await readAgentConfig()
     const maxSteps = agentConfig.maxSteps
diff --git a/src/ai-providers/types.ts b/src/ai-providers/types.ts
index 59d75bb7..873c92ee 100644
--- a/src/ai-providers/types.ts
+++ b/src/ai-providers/types.ts
@@ -1,6 +1,7 @@
 import type { ISessionStore, SessionEntry } from '../core/session.js'
 import type { CompactionConfig, CompactionResult } from '../core/compaction.js'
 import type { MediaAttachment } from '../core/types.js'
+import type { ResolvedProfile } from '../core/config.js'
 
 // ==================== Provider Events ====================
 
@@ -30,9 +31,8 @@ export interface GenerateOpts {
   /** Max history entries to include (text providers only). */
   maxHistoryEntries?: number
   disabledTools?: string[]
-  vercelAiSdk?: { provider: string; model: string; baseUrl?: string; apiKey?: string }
-  agentSdk?: { model?: string; apiKey?: string; baseUrl?: string }
-  codex?: { model?: string; baseUrl?: string; apiKey?: string; loginMethod?: 'api-key' | 'codex-oauth' }
+  /** Resolved profile — contains model, apiKey, baseUrl, etc. */
+  profile?: ResolvedProfile
 }
 
 // ==================== AIProvider ====================
diff --git a/src/ai-providers/vercel-ai-sdk/model-factory.ts b/src/ai-providers/vercel-ai-sdk/model-factory.ts
index 0e42c3f2..dd473c77 100644
--- a/src/ai-providers/vercel-ai-sdk/model-factory.ts
+++ b/src/ai-providers/vercel-ai-sdk/model-factory.ts
@@ -1,13 +1,11 @@
 /**
- * Model factory — creates Vercel AI SDK LanguageModel instances from config.
+ * Model factory — creates Vercel AI SDK LanguageModel instances from a resolved profile.
  *
- * Reads ai-provider-manager.json from disk on each call so that model
- * changes take effect without a restart.  Uses dynamic imports so unused
- * provider packages don't prevent startup.
+ * Uses dynamic imports so unused provider packages don't prevent startup.
  */
 
 import type { LanguageModel } from 'ai'
-import { readAIProviderConfig } from '../../core/config.js'
+import type { ResolvedProfile } from '../../core/config.js'
 
 /** Result includes the model plus a cache key for change detection. */
 export interface ModelFromConfig {
@@ -16,42 +14,27 @@ export interface ModelFromConfig {
   key: string
 }
 
-/** Per-request model override (e.g. from a sub-channel's vercelAiSdk config). */
-export interface ModelOverride {
-  provider: string
-  model: string
-  baseUrl?: string
-  apiKey?: string
-}
-
-export async function createModelFromConfig(override?: ModelOverride): Promise<ModelFromConfig> {
-  // Resolve effective values: override takes precedence over global config
-  const config = await readAIProviderConfig()
-  const p = override?.provider ?? config.provider
-  const m = override?.model ?? config.model
-  const url = override?.baseUrl ?? config.baseUrl
+export async function createModelFromProfile(profile: ResolvedProfile): Promise<ModelFromConfig> {
+  const p = profile.provider ?? 'anthropic'
+  const m = profile.model
+  const url = profile.baseUrl
+  const apiKey = profile.apiKey
   const key = `${p}:${m}:${url ?? ''}`
 
-  // Resolve API key: override.apiKey > global config.apiKeys[provider]
-  const resolveApiKey = (provider: string) => {
-    if (override?.apiKey) return override.apiKey
-    return (config.apiKeys as Record<string, string | undefined>)[provider] || undefined
-  }
-
   switch (p) {
     case 'anthropic': {
       const { createAnthropic } = await import('@ai-sdk/anthropic')
-      const client = createAnthropic({ apiKey: resolveApiKey('anthropic'), baseURL: url || undefined })
+      const client = createAnthropic({ apiKey: apiKey || undefined, baseURL: url || undefined })
       return { model: client(m), key }
     }
     case 'openai': {
       const { createOpenAI } = await import('@ai-sdk/openai')
-      const client = createOpenAI({ apiKey: resolveApiKey('openai'), baseURL: url || undefined })
+      const client = createOpenAI({ apiKey: apiKey || undefined, baseURL: url || undefined })
       return { model: client(m), key }
     }
     case 'google': {
       const { createGoogleGenerativeAI } = await import('@ai-sdk/google')
-      const client = createGoogleGenerativeAI({ apiKey: resolveApiKey('google'), baseURL: url || undefined })
+      const client = createGoogleGenerativeAI({ apiKey: apiKey || undefined, baseURL: url || undefined })
       return { model: client(m), key }
     }
     default:
diff --git a/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts b/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts
index 0d3b24b8..9b77839f 100644
--- a/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts
+++ b/src/ai-providers/vercel-ai-sdk/vercel-provider.spec.ts
@@ -2,7 +2,11 @@ import { describe, it, expect, vi, beforeEach } from 'vitest'
 import { VercelAIProvider } from './vercel-provider.js'
 
 vi.mock('./model-factory.js', () => ({
-  createModelFromConfig: vi.fn(),
+  createModelFromProfile: vi.fn(),
+}))
+
+vi.mock('../../core/config.js', () => ({
+  resolveProfile: vi.fn().mockResolvedValue({ backend: 'vercel-ai-sdk', label: 'Test', model: 'mock-model', provider: 'anthropic' }),
 }))
 
 vi.mock('./agent.js', () => ({
@@ -14,10 +18,10 @@ vi.mock('../../core/media.js', () => ({
   extractMediaFromToolOutput: vi.fn().mockReturnValue([]),
 }))
 
-import { createModelFromConfig } from './model-factory.js'
+import { createModelFromProfile } from './model-factory.js'
 import { generateText } from './agent.js'
 
-const mockCreateModelFromConfig = vi.mocked(createModelFromConfig)
+const mockCreateModelFromProfile = vi.mocked(createModelFromProfile)
 const mockGenerateText = vi.mocked(generateText)
 
 // ==================== Helpers ====================
@@ -32,7 +36,7 @@ function makeProvider(overrides?: { getTools?: () => Promise<Record<string, any>
 describe('VercelAIProvider — ask()', () => {
   beforeEach(() => {
     vi.clearAllMocks()
-    mockCreateModelFromConfig.mockResolvedValue({ model: {} as any, key: 'gpt-4o' })
+    mockCreateModelFromProfile.mockResolvedValue({ model: {} as any, key: 'gpt-4o' })
     mockGenerateText.mockResolvedValue({ text: 'ok', steps: [] } as any)
   })
 
@@ -68,7 +72,7 @@ describe('VercelAIProvider — ask()', () => {
 describe('VercelAIProvider — generate() tool filtering', () => {
   beforeEach(() => {
     vi.clearAllMocks()
-    mockCreateModelFromConfig.mockResolvedValue({ model: {} as any, key: 'gpt-4o' })
+    mockCreateModelFromProfile.mockResolvedValue({ model: {} as any, key: 'gpt-4o' })
     mockGenerateText.mockResolvedValue({ text: 'ok', steps: [] } as any)
   })
 
@@ -88,14 +92,15 @@ describe('VercelAIProvider — generate() tool filtering', () => {
     expect(toolNames).toContain('toolC')
   })
 
-  it('passes modelOverride to createModelFromConfig', async () => {
+  it('passes profile to createModelFromProfile', async () => {
     const provider = makeProvider()
+    const profile = { backend: 'vercel-ai-sdk' as const, label: 'Test', model: 'claude-3-7', provider: 'anthropic' }
 
-    for await (const _ of provider.generate([], 'test', { vercelAiSdk: { modelId: 'claude-3-7' } as any })) {
+    for await (const _ of provider.generate([], 'test', { profile })) {
       // drain
     }
 
-    expect(mockCreateModelFromConfig).toHaveBeenCalledWith({ modelId: 'claude-3-7' })
+    expect(mockCreateModelFromProfile).toHaveBeenCalledWith(profile)
   })
 })
 
@@ -104,7 +109,7 @@ describe('VercelAIProvider — generate() tool filtering', () => {
 describe('VercelAIProvider — generate() events', () => {
   beforeEach(() => {
     vi.clearAllMocks()
-    mockCreateModelFromConfig.mockResolvedValue({ model: {} as any, key: 'gpt-4o' })
+    mockCreateModelFromProfile.mockResolvedValue({ model: {} as any, key: 'gpt-4o' })
   })
 
   it('yields done event with text from result', async () => {
diff --git a/src/ai-providers/vercel-ai-sdk/vercel-provider.ts b/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
index eb76a8fa..b09da3be 100644
--- a/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
+++ b/src/ai-providers/vercel-ai-sdk/vercel-provider.ts
@@ -1,8 +1,7 @@
 /**
  * VercelAIProvider — GenerateProvider backed by Vercel AI SDK's generateText.
  *
- * The model is lazily created from config on each call. When ai-provider.json
- * changes on disk, the next request picks up the new model automatically.
+ * The model is lazily created from the resolved profile on each call.
  * Instructions (persona + brain state) are also re-read per request.
  */
 
@@ -10,9 +9,11 @@ import type { ModelMessage, Tool } from 'ai'
 import type { ProviderResult, ProviderEvent, AIProvider, GenerateOpts } from '../types.js'
 import type { SessionEntry } from '../../core/session.js'
 import type { MediaAttachment } from '../../core/types.js'
+import type { ResolvedProfile } from '../../core/config.js'
+import { resolveProfile } from '../../core/config.js'
 import { toModelMessages } from '../../core/session.js'
 import { extractMediaFromToolOutput } from '../../core/media.js'
-import { createModelFromConfig, type ModelOverride } from './model-factory.js'
+import { createModelFromProfile } from './model-factory.js'
 import { generateText, stepCountIs } from './agent.js'
 import { createChannel } from '../../core/async-channel.js'
 
@@ -26,9 +27,11 @@ export class VercelAIProvider implements AIProvider {
   ) {}
 
   /** Resolve model, tools, and instructions for a single request. */
-  private async resolve(disabledTools?: string[], modelOverride?: ModelOverride) {
+  private async resolve(disabledTools?: string[], profile?: ResolvedProfile) {
+    // If no profile provided (e.g. ask()), resolve the active one
+    const effectiveProfile = profile ?? await resolveProfile()
     const [{ model }, allTools, instructions] = await Promise.all([
-      createModelFromConfig(modelOverride),
+      createModelFromProfile(effectiveProfile),
       this.getTools(),
       this.getInstructions(),
     ])
@@ -61,7 +64,7 @@ export class VercelAIProvider implements AIProvider {
   }
 
   async *generate(entries: SessionEntry[], _prompt: string, opts?: GenerateOpts): AsyncGenerator<ProviderEvent> {
-    const { model, tools, instructions } = await this.resolve(opts?.disabledTools, opts?.vercelAiSdk)
+    const { model, tools, instructions } = await this.resolve(opts?.disabledTools, opts?.profile)
     const messages = toModelMessages(entries)
 
     const channel = createChannel<ProviderEvent>()
diff --git a/src/connectors/telegram/telegram-plugin.ts b/src/connectors/telegram/telegram-plugin.ts
index a6499b90..1293c675 100644
--- a/src/connectors/telegram/telegram-plugin.ts
+++ b/src/connectors/telegram/telegram-plugin.ts
@@ -10,18 +10,16 @@ import { askAgentSdk } from '../../ai-providers/agent-sdk/query.js'
 import type { AgentSdkConfig } from '../../ai-providers/agent-sdk/query.js'
 import { SessionStore } from '../../core/session'
 import { forceCompact } from '../../core/compaction'
-import { readAIBackend, writeAIBackend, readConnectorsConfig, type AIBackend } from '../../core/config'
+import { readAIProviderConfig, setActiveProfile, readConnectorsConfig } from '../../core/config'
 import type { ConnectorCenter } from '../../core/connector-center.js'
 import { TelegramConnector, splitMessage, MAX_MESSAGE_LENGTH } from './telegram-connector.js'
 import type { AccountManager } from '../../domain/trading/index.js'
 import type { Operation } from '../../domain/trading/git/types.js'
 import { getOperationSymbol } from '../../domain/trading/git/types.js'
 
-const BACKEND_LABELS: Record<AIBackend, string> = {
-  'claude-code': 'Claude Code',
-  'vercel-ai-sdk': 'Vercel AI SDK',
-  'agent-sdk': 'Agent SDK',
-  'codex': 'Codex',
+/** Build a display label for a profile. */
+function profileLabel(slug: string, profile: { label: string; backend: string; model: string }): string {
+  return `${profile.label} (${profile.model})`
 }
 
 export class TelegramPlugin implements Plugin {
@@ -89,8 +87,10 @@ export class TelegramPlugin implements Plugin {
 
     // ── Commands ──
     bot.command('status', async (ctx) => {
-      const aiConfig = await readAIBackend()
-      await this.sendReply(ctx.chat.id, `Engine is running. Provider: ${BACKEND_LABELS[aiConfig.backend]}`)
+      const aiConfig = await readAIProviderConfig()
+      const profile = aiConfig.profiles[aiConfig.activeProfile]
+      const label = profile ? profileLabel(aiConfig.activeProfile, profile) : aiConfig.activeProfile
+      await this.sendReply(ctx.chat.id, `Engine is running. Profile: ${label}`)
     })
 
     bot.command('settings', async (ctx) => {
@@ -115,21 +115,22 @@ export class TelegramPlugin implements Plugin {
     bot.on('callback_query:data', async (ctx) => {
       const data = ctx.callbackQuery.data
       try {
-        if (data.startsWith('provider:')) {
-          const backend = data.slice('provider:'.length) as AIBackend
-          await writeAIBackend(backend)
-          await ctx.answerCallbackQuery({ text: `Switched to ${BACKEND_LABELS[backend]}` })
+        if (data.startsWith('profile:')) {
+          const slug = data.slice('profile:'.length)
+          await setActiveProfile(slug)
+          const config = await readAIProviderConfig()
+          const profile = config.profiles[slug]
+          const label = profile ? profileLabel(slug, profile) : slug
+          await ctx.answerCallbackQuery({ text: `Switched to ${label}` })
 
           // Edit the original settings message in-place
-          const ccLabel = backend === 'claude-code' ? '> Claude Code' : 'Claude Code'
-          const aiLabel = backend === 'vercel-ai-sdk' ? '> Vercel AI SDK' : 'Vercel AI SDK'
-          const sdkLabel = backend === 'agent-sdk' ? '> Agent SDK' : 'Agent SDK'
           const keyboard = new InlineKeyboard()
-            .text(ccLabel, 'provider:claude-code')
-            .text(aiLabel, 'provider:vercel-ai-sdk')
-            .text(sdkLabel, 'provider:agent-sdk')
+          for (const [s, p] of Object.entries(config.profiles)) {
+            const prefix = s === slug ? '> ' : ''
+            keyboard.text(`${prefix}${p.label}`, `profile:${s}`)
+          }
           await ctx.editMessageText(
-            `Current provider: ${BACKEND_LABELS[backend]}\n\nChoose default AI provider:`,
+            `Current profile: ${label}\n\nChoose AI profile:`,
             { reply_markup: keyboard },
           )
         } else if (data.startsWith('trading:')) {
@@ -218,8 +219,8 @@ export class TelegramPlugin implements Plugin {
 
     // ── Initialize and get bot info ──
     await bot.init()
-    const aiConfig = await readAIBackend()
-    console.log(`telegram plugin: connected as @${bot.botInfo.username} (backend: ${aiConfig.backend})`)
+    const initConfig = await readAIProviderConfig()
+    console.log(`telegram plugin: connected as @${bot.botInfo.username} (profile: ${initConfig.activeProfile})`)
 
     // ── Register connector for outbound delivery (heartbeat / cron responses) ──
     if (this.config.allowedChatIds.length > 0) {
@@ -337,19 +338,19 @@ export class TelegramPlugin implements Plugin {
   }
 
   private async sendSettingsMenu(chatId: number) {
-    const aiConfig = await readAIBackend()
-    const ccLabel = aiConfig.backend === 'claude-code' ? '> Claude Code' : 'Claude Code'
-    const aiLabel = aiConfig.backend === 'vercel-ai-sdk' ? '> Vercel AI SDK' : 'Vercel AI SDK'
-    const sdkLabel = aiConfig.backend === 'agent-sdk' ? '> Agent SDK' : 'Agent SDK'
+    const config = await readAIProviderConfig()
+    const activeProfile = config.profiles[config.activeProfile]
+    const activeLabel = activeProfile ? profileLabel(config.activeProfile, activeProfile) : config.activeProfile
 
     const keyboard = new InlineKeyboard()
-      .text(ccLabel, 'provider:claude-code')
-      .text(aiLabel, 'provider:vercel-ai-sdk')
-      .text(sdkLabel, 'provider:agent-sdk')
+    for (const [slug, profile] of Object.entries(config.profiles)) {
+      const prefix = slug === config.activeProfile ? '> ' : ''
+      keyboard.text(`${prefix}${profile.label}`, `profile:${slug}`)
+    }
 
     await this.bot!.api.sendMessage(
       chatId,
-      `Current provider: ${BACKEND_LABELS[aiConfig.backend]}\n\nChoose default AI provider:`,
+      `Current profile: ${activeLabel}\n\nChoose AI profile:`,
       { reply_markup: keyboard },
     )
   }
diff --git a/src/connectors/web/__tests__/chat-streaming.spec.ts b/src/connectors/web/__tests__/chat-streaming.spec.ts
index 70e909f9..fc4358aa 100644
--- a/src/connectors/web/__tests__/chat-streaming.spec.ts
+++ b/src/connectors/web/__tests__/chat-streaming.spec.ts
@@ -25,6 +25,11 @@ import type { SSEClient } from '../routes/chat.js'
 
 // ==================== Module Mocks ====================
 
+vi.mock('../../../core/config.js', () => ({
+  resolveProfile: vi.fn().mockResolvedValue({ backend: 'vercel-ai-sdk', label: 'Test', model: 'mock', provider: 'anthropic' }),
+  readAgentConfig: vi.fn().mockResolvedValue({ maxSteps: 20, evolutionMode: false, claudeCode: { disallowedTools: [], maxTurns: 20 } }),
+}))
+
 vi.mock('../../../core/compaction.js', async (importOriginal) => {
   const actual = await importOriginal<typeof import('../../../core/compaction.js')>()
   return {
diff --git a/src/connectors/web/routes/channels.ts b/src/connectors/web/routes/channels.ts
index 7d1b40bc..5f1fa86b 100644
--- a/src/connectors/web/routes/channels.ts
+++ b/src/connectors/web/routes/channels.ts
@@ -29,10 +29,7 @@ export function createChannelsRoutes({ sessions, sseByChannel }: ChannelsDeps) {
       id?: string
       label?: string
       systemPrompt?: string
-      provider?: string
-      vercelAiSdk?: { provider: string; model: string; baseUrl?: string; apiKey?: string }
-      agentSdk?: { model?: string; apiKey?: string; baseUrl?: string }
-      codex?: { model?: string; baseUrl?: string; apiKey?: string; loginMethod?: 'api-key' | 'codex-oauth' }
+      profile?: string
       disabledTools?: string[]
     }
 
@@ -55,14 +52,7 @@ export function createChannelsRoutes({ sessions, sseByChannel }: ChannelsDeps) {
       id: body.id,
       label: body.label.trim(),
       ...(body.systemPrompt ? { systemPrompt: body.systemPrompt } : {}),
-      ...(body.provider === 'claude-code' || body.provider === 'vercel-ai-sdk' || body.provider === 'agent-sdk' || body.provider === 'codex'
-        ? { provider: body.provider }
-        : {}),
-      ...(body.vercelAiSdk?.provider && body.vercelAiSdk?.model
-        ? { vercelAiSdk: body.vercelAiSdk }
-        : {}),
-      ...(body.agentSdk ? { agentSdk: body.agentSdk } : {}),
-      ...(body.codex ? { codex: body.codex } : {}),
+      ...(body.profile ? { profile: body.profile } : {}),
       ...(body.disabledTools?.length ? { disabledTools: body.disabledTools } : {}),
     }
 
@@ -85,10 +75,7 @@ export function createChannelsRoutes({ sessions, sseByChannel }: ChannelsDeps) {
     const body = await c.req.json() as {
       label?: string
       systemPrompt?: string
-      provider?: string
-      vercelAiSdk?: { provider: string; model: string; baseUrl?: string; apiKey?: string } | null
-      agentSdk?: { model?: string; apiKey?: string; baseUrl?: string } | null
-      codex?: { model?: string; baseUrl?: string; apiKey?: string; loginMethod?: 'api-key' | 'codex-oauth' } | null
+      profile?: string | null
       disabledTools?: string[]
     }
 
@@ -100,20 +87,7 @@ export function createChannelsRoutes({ sessions, sseByChannel }: ChannelsDeps) {
       ...existing[idx],
       ...(body.label !== undefined ? { label: body.label } : {}),
       ...(body.systemPrompt !== undefined ? { systemPrompt: body.systemPrompt || undefined } : {}),
-      ...(body.provider === 'claude-code' || body.provider === 'vercel-ai-sdk' || body.provider === 'agent-sdk' || body.provider === 'codex'
-        ? { provider: body.provider }
-        : body.provider === null || body.provider === ''
-          ? { provider: undefined }
-          : {}),
-      ...(body.vercelAiSdk !== undefined
-        ? { vercelAiSdk: body.vercelAiSdk?.provider && body.vercelAiSdk?.model ? body.vercelAiSdk : undefined }
-        : {}),
-      ...(body.agentSdk !== undefined
-        ? { agentSdk: body.agentSdk ?? undefined }
-        : {}),
-      ...(body.codex !== undefined
-        ? { codex: body.codex ?? undefined }
-        : {}),
+      ...(body.profile !== undefined ? { profile: body.profile || undefined } : {}),
       ...(body.disabledTools !== undefined ? { disabledTools: body.disabledTools?.length ? body.disabledTools : undefined } : {}),
     }
     existing[idx] = updated
diff --git a/src/connectors/web/routes/chat.ts b/src/connectors/web/routes/chat.ts
index 354894ba..3b9fa18d 100644
--- a/src/connectors/web/routes/chat.ts
+++ b/src/connectors/web/routes/chat.ts
@@ -43,10 +43,7 @@ export function createChatRoutes({ ctx, sessions, sseByChannel }: ChatDeps) {
       if (channel) {
         if (channel.systemPrompt) opts.systemPrompt = channel.systemPrompt
         if (channel.disabledTools?.length) opts.disabledTools = channel.disabledTools
-        if (channel.provider) opts.provider = channel.provider
-        if (channel.vercelAiSdk) opts.vercelAiSdk = channel.vercelAiSdk
-        if (channel.agentSdk) opts.agentSdk = channel.agentSdk
-        if (channel.codex) opts.codex = channel.codex
+        if (channel.profile) opts.profileSlug = channel.profile
       }
     }
 
diff --git a/src/connectors/web/routes/config.ts b/src/connectors/web/routes/config.ts
index 26036109..b221b685 100644
--- a/src/connectors/web/routes/config.ts
+++ b/src/connectors/web/routes/config.ts
@@ -1,12 +1,16 @@
 import { Hono } from 'hono'
-import { loadConfig, writeConfigSection, readAIProviderConfig, validSections, writeAIBackend, type ConfigSection, type AIBackend } from '../../../core/config.js'
+import {
+  loadConfig, writeConfigSection, readAIProviderConfig, validSections,
+  writeProfile, deleteProfile, setActiveProfile, writeApiKeys,
+  profileSchema, type ConfigSection, type Profile,
+} from '../../../core/config.js'
 import type { EngineContext } from '../../../core/types.js'
 
 interface ConfigRouteOpts {
   onConnectorsChange?: () => Promise<void>
 }
 
-/** Config routes: GET /, PUT /ai-provider, PUT /:section, GET /api-keys/status */
+/** Config routes: GET /, PUT /:section, profile CRUD, api-keys */
 export function createConfigRoutes(opts?: ConfigRouteOpts) {
   const app = new Hono()
 
@@ -19,33 +23,32 @@ export function createConfigRoutes(opts?: ConfigRouteOpts) {
     }
   })
 
-  app.put('/ai-provider', async (c) => {
+  // ==================== Profile CRUD ====================
+
+  /** GET /profiles — list all profiles */
+  app.get('/profiles', async (c) => {
     try {
-      const body = await c.req.json<{ backend?: string }>()
-      const backend = body.backend
-      if (backend !== 'claude-code' && backend !== 'vercel-ai-sdk' && backend !== 'agent-sdk' && backend !== 'codex') {
-        return c.json({ error: 'Invalid backend. Must be "claude-code", "vercel-ai-sdk", "agent-sdk", or "codex".' }, 400)
-      }
-      await writeAIBackend(backend as AIBackend)
-      return c.json({ backend })
+      const config = await readAIProviderConfig()
+      return c.json({ profiles: config.profiles, activeProfile: config.activeProfile })
     } catch (err) {
       return c.json({ error: String(err) }, 500)
     }
   })
 
-  app.put('/:section', async (c) => {
+  /** POST /profiles — create a new profile */
+  app.post('/profiles', async (c) => {
     try {
-      const section = c.req.param('section') as ConfigSection
-      if (!validSections.includes(section)) {
-        return c.json({ error: `Invalid section "${section}". Valid: ${validSections.join(', ')}` }, 400)
+      const body = await c.req.json<{ slug: string; profile: Profile }>()
+      if (!body.slug || !/^[a-z0-9-]+$/.test(body.slug)) {
+        return c.json({ error: 'slug must be lowercase alphanumeric with hyphens' }, 400)
       }
-      const body = await c.req.json()
-      const validated = await writeConfigSection(section, body)
-      // Hot-reload connectors / OpenBB server when their config changes
-      if (section === 'connectors' || section === 'marketData') {
-        await opts?.onConnectorsChange?.()
+      const config = await readAIProviderConfig()
+      if (config.profiles[body.slug]) {
+        return c.json({ error: 'profile slug already exists' }, 409)
       }
-      return c.json(validated)
+      const validated = profileSchema.parse(body.profile)
+      await writeProfile(body.slug, validated)
+      return c.json({ slug: body.slug, profile: validated }, 201)
     } catch (err) {
       if (err instanceof Error && err.name === 'ZodError') {
         return c.json({ error: 'Validation failed', details: JSON.parse(err.message) }, 400)
@@ -54,6 +57,57 @@ export function createConfigRoutes(opts?: ConfigRouteOpts) {
     }
   })
 
+  /** PUT /profiles/:slug — update a profile */
+  app.put('/profiles/:slug', async (c) => {
+    try {
+      const slug = c.req.param('slug')
+      const body = await c.req.json<Profile>()
+      const validated = profileSchema.parse(body)
+      await writeProfile(slug, validated)
+      return c.json({ slug, profile: validated })
+    } catch (err) {
+      if (err instanceof Error && err.name === 'ZodError') {
+        return c.json({ error: 'Validation failed', details: JSON.parse(err.message) }, 400)
+      }
+      return c.json({ error: String(err) }, 500)
+    }
+  })
+
+  /** DELETE /profiles/:slug — delete a profile */
+  app.delete('/profiles/:slug', async (c) => {
+    try {
+      const slug = c.req.param('slug')
+      await deleteProfile(slug)
+      return c.json({ success: true })
+    } catch (err) {
+      return c.json({ error: String(err) }, 400)
+    }
+  })
+
+  /** PUT /active-profile — set the active profile */
+  app.put('/active-profile', async (c) => {
+    try {
+      const { slug } = await c.req.json<{ slug: string }>()
+      await setActiveProfile(slug)
+      return c.json({ activeProfile: slug })
+    } catch (err) {
+      return c.json({ error: String(err) }, 400)
+    }
+  })
+
+  // ==================== API Keys ====================
+
+  /** PUT /api-keys — update global API keys */
+  app.put('/api-keys', async (c) => {
+    try {
+      const body = await c.req.json<{ anthropic?: string; openai?: string; google?: string }>()
+      await writeApiKeys(body)
+      return c.json({ success: true })
+    } catch (err) {
+      return c.json({ error: String(err) }, 500)
+    }
+  })
+
   app.get('/api-keys/status', async (c) => {
     try {
       const config = await readAIProviderConfig()
@@ -67,6 +121,29 @@ export function createConfigRoutes(opts?: ConfigRouteOpts) {
     }
   })
 
+  // ==================== Generic Section Writer ====================
+
+  app.put('/:section', async (c) => {
+    try {
+      const section = c.req.param('section') as ConfigSection
+      if (!validSections.includes(section)) {
+        return c.json({ error: `Invalid section "${section}". Valid: ${validSections.join(', ')}` }, 400)
+      }
+      const body = await c.req.json()
+      const validated = await writeConfigSection(section, body)
+      // Hot-reload connectors / OpenBB server when their config changes
+      if (section === 'connectors' || section === 'marketData') {
+        await opts?.onConnectorsChange?.()
+      }
+      return c.json(validated)
+    } catch (err) {
+      if (err instanceof Error && err.name === 'ZodError') {
+        return c.json({ error: 'Validation failed', details: JSON.parse(err.message) }, 400)
+      }
+      return c.json({ error: String(err) }, 500)
+    }
+  })
+
   return app
 }
 
diff --git a/src/core/__tests__/pipeline/helpers.ts b/src/core/__tests__/pipeline/helpers.ts
index cb1806bb..d8566428 100644
--- a/src/core/__tests__/pipeline/helpers.ts
+++ b/src/core/__tests__/pipeline/helpers.ts
@@ -6,10 +6,17 @@
  * imported from their respective modules (MemorySessionStore, MockConnector).
  */
 
+import { vi } from 'vitest'
 import { AgentCenter } from '../../agent-center.js'
 import { GenerateRouter, StreamableResult, type ProviderEvent } from '../../ai-provider-manager.js'
 import { DEFAULT_COMPACTION_CONFIG } from '../../compaction.js'
 
+// Mock resolveProfile so GenerateRouter.resolve() works without disk I/O
+vi.mock('../../config.js', () => ({
+  resolveProfile: vi.fn().mockResolvedValue({ backend: 'vercel-ai-sdk', label: 'Test', model: 'mock', provider: 'anthropic' }),
+  readAgentConfig: vi.fn().mockResolvedValue({ maxSteps: 20, evolutionMode: false, claudeCode: { disallowedTools: [], maxTurns: 20 } }),
+}))
+
 // Re-export test doubles for convenience
 export { MemorySessionStore } from '../../session.js'
 export type { SessionEntry, ContentBlock } from '../../session.js'
diff --git a/src/core/__tests__/pipeline/streaming.spec.ts b/src/core/__tests__/pipeline/streaming.spec.ts
index 60057764..0bc582d0 100644
--- a/src/core/__tests__/pipeline/streaming.spec.ts
+++ b/src/core/__tests__/pipeline/streaming.spec.ts
@@ -19,6 +19,11 @@ import {
 
 // ==================== Module Mocks ====================
 
+vi.mock('../../config.js', () => ({
+  resolveProfile: vi.fn().mockResolvedValue({ backend: 'vercel-ai-sdk', label: 'Test', model: 'mock', provider: 'anthropic' }),
+  readAgentConfig: vi.fn().mockResolvedValue({ maxSteps: 20, evolutionMode: false, claudeCode: { disallowedTools: [], maxTurns: 20 } }),
+}))
+
 vi.mock('../../compaction.js', async (importOriginal) => {
   const actual = await importOriginal<typeof import('../../compaction.js')>()
   return {
diff --git a/src/core/agent-center.spec.ts b/src/core/agent-center.spec.ts
index 61f41abd..c8406c38 100644
--- a/src/core/agent-center.spec.ts
+++ b/src/core/agent-center.spec.ts
@@ -5,7 +5,7 @@ import { AgentCenter } from './agent-center.js'
 import { GenerateRouter } from './ai-provider-manager.js'
 import { DEFAULT_COMPACTION_CONFIG, type CompactionConfig } from './compaction.js'
 import { VercelAIProvider } from '../ai-providers/vercel-ai-sdk/vercel-provider.js'
-import { createModelFromConfig } from '../ai-providers/vercel-ai-sdk/model-factory.js'
+import { createModelFromProfile } from '../ai-providers/vercel-ai-sdk/model-factory.js'
 import { MemorySessionStore, type SessionEntry } from './session.js'
 
 // ==================== Helpers ====================
@@ -42,7 +42,7 @@ function makeAgentCenter(overrides: MakeAgentCenterOpts = {}): AgentCenter {
   const maxSteps = overrides.maxSteps ?? 1
   const compaction = overrides.compaction ?? DEFAULT_COMPACTION_CONFIG
 
-  vi.mocked(createModelFromConfig).mockResolvedValue({ model, key: 'test:mock-model' })
+  vi.mocked(createModelFromProfile).mockResolvedValue({ model, key: 'test:mock-model' })
   const provider = new VercelAIProvider(async () => tools, async () => instructions, maxSteps)
   const router = new GenerateRouter(provider, null)
 
@@ -52,7 +52,12 @@ function makeAgentCenter(overrides: MakeAgentCenterOpts = {}): AgentCenter {
 // ==================== Mock model-factory ====================
 
 vi.mock('../ai-providers/vercel-ai-sdk/model-factory.js', () => ({
-  createModelFromConfig: vi.fn(),
+  createModelFromProfile: vi.fn(),
+}))
+
+vi.mock('./config.js', () => ({
+  resolveProfile: vi.fn().mockResolvedValue({ backend: 'vercel-ai-sdk', label: 'Test', model: 'mock-model', provider: 'anthropic' }),
+  readAgentConfig: vi.fn().mockResolvedValue({ maxSteps: 20, evolutionMode: false, claudeCode: { disallowedTools: [], maxTurns: 20 } }),
 }))
 
 // ==================== Mock compaction ====================
diff --git a/src/core/agent-center.ts b/src/core/agent-center.ts
index 5c6afc6f..bd2830f3 100644
--- a/src/core/agent-center.ts
+++ b/src/core/agent-center.ts
@@ -73,8 +73,8 @@ export class AgentCenter {
     // 1. Append user message to session
     await session.appendUser(prompt, 'human')
 
-    // 2. Resolve provider (may be overridden per-request)
-    const provider = await this.router.resolve(opts?.provider)
+    // 2. Resolve provider + profile (may be overridden per-request via profileSlug)
+    const { provider, profile } = await this.router.resolve(opts?.profileSlug)
 
     // 3. Compact if needed (provider can override with custom strategy)
     const compactionResult = provider.compact
@@ -94,9 +94,7 @@ export class AgentCenter {
       historyPreamble: opts?.historyPreamble ?? this.defaultPreamble,
       maxHistoryEntries: opts?.maxHistoryEntries ?? this.defaultMaxHistory,
       disabledTools: opts?.disabledTools,
-      vercelAiSdk: opts?.vercelAiSdk,
-      agentSdk: opts?.agentSdk,
-      codex: opts?.codex,
+      profile,
     }
     const source = provider.generate(entries, prompt, genOpts)
 
diff --git a/src/core/ai-provider-manager.spec.ts b/src/core/ai-provider-manager.spec.ts
index f611d5b4..ed665a4b 100644
--- a/src/core/ai-provider-manager.spec.ts
+++ b/src/core/ai-provider-manager.spec.ts
@@ -129,6 +129,13 @@ describe('StreamableResult', () => {
 
 // ==================== GenerateRouter ====================
 
+vi.mock('./config.js', () => ({
+  resolveProfile: vi.fn(),
+}))
+
+import { resolveProfile } from './config.js'
+const mockResolveProfile = vi.mocked(resolveProfile)
+
 describe('GenerateRouter', () => {
   function makeProvider(tag: AIProvider['providerTag']): AIProvider {
     return {
@@ -138,56 +145,49 @@ describe('GenerateRouter', () => {
     }
   }
 
-  it('should resolve to vercel when no override and config fallback', async () => {
-    const vercel = makeProvider('vercel-ai')
-    const router = new GenerateRouter(vercel, null)
-
-    // Without override, reads config — agentSdk is null so falls back to vercel
-    const provider = await router.resolve()
-    expect(provider).toBe(vercel)
-  })
-
-  it('should resolve override claude-code as alias for agent-sdk', async () => {
+  it('should resolve profile and pick matching provider', async () => {
     const vercel = makeProvider('vercel-ai')
     const agentSdk = makeProvider('agent-sdk')
     const router = new GenerateRouter(vercel, agentSdk)
 
-    const provider = await router.resolve('claude-code')
+    mockResolveProfile.mockResolvedValue({ backend: 'agent-sdk', label: 'Claude', model: 'claude-sonnet-4-6' })
+    const { provider } = await router.resolve('claude-main')
     expect(provider).toBe(agentSdk)
   })
 
-  it('should resolve override agent-sdk when available', async () => {
+  it('should resolve active profile when no slug given', async () => {
     const vercel = makeProvider('vercel-ai')
-    const agentSdk = makeProvider('agent-sdk')
-    const router = new GenerateRouter(vercel, agentSdk)
+    const router = new GenerateRouter(vercel, null)
 
-    const provider = await router.resolve('agent-sdk')
-    expect(provider).toBe(agentSdk)
+    mockResolveProfile.mockResolvedValue({ backend: 'vercel-ai-sdk', label: 'Vercel', model: 'claude-sonnet-4-6', provider: 'anthropic' })
+    const { provider } = await router.resolve()
+    expect(provider).toBe(vercel)
   })
 
-  it('should fallback to vercel when override claude-code but not available', async () => {
+  it('should throw when backend has no registered provider', async () => {
     const vercel = makeProvider('vercel-ai')
-    const router = new GenerateRouter(vercel, null)
+    const router = new GenerateRouter(vercel, null) // no agent-sdk
 
-    // Override requests claude-code but it's null — falls through to config read
-    const provider = await router.resolve('claude-code')
-    // Config read will happen, but since claudeCode is null, falls to vercel
-    expect(provider).toBe(vercel)
+    mockResolveProfile.mockResolvedValue({ backend: 'agent-sdk', label: 'Claude', model: 'x' })
+    await expect(router.resolve('test')).rejects.toThrow('No provider registered for backend')
   })
 
-  it('should resolve override vercel-ai-sdk directly', async () => {
+  it('should resolve codex provider', async () => {
     const vercel = makeProvider('vercel-ai')
-    const cc = makeProvider('claude-code')
-    const router = new GenerateRouter(vercel, cc)
+    const codex = makeProvider('codex')
+    const router = new GenerateRouter(vercel, null, codex)
 
-    const provider = await router.resolve('vercel-ai-sdk')
-    expect(provider).toBe(vercel)
+    mockResolveProfile.mockResolvedValue({ backend: 'codex', label: 'GPT', model: 'gpt-5.4' })
+    const { provider, profile } = await router.resolve('gpt-main')
+    expect(provider).toBe(codex)
+    expect(profile.model).toBe('gpt-5.4')
   })
 
-  it('should delegate ask to resolved provider', async () => {
+  it('should delegate ask to active profile provider', async () => {
     const vercel = makeProvider('vercel-ai')
     const router = new GenerateRouter(vercel, null)
 
+    mockResolveProfile.mockResolvedValue({ backend: 'vercel-ai-sdk', label: 'V', model: 'x', provider: 'anthropic' })
     const result = await router.ask('test prompt')
     expect(result.text).toBe('from-vercel-ai')
     expect(vercel.ask).toHaveBeenCalledWith('test prompt')
diff --git a/src/core/ai-provider-manager.ts b/src/core/ai-provider-manager.ts
index ee3342f3..b37f14fd 100644
--- a/src/core/ai-provider-manager.ts
+++ b/src/core/ai-provider-manager.ts
@@ -6,7 +6,8 @@
  * core infrastructure that orchestrates providers.
  */
 
-import { readAIProviderConfig } from './config.js'
+import { resolveProfile } from './config.js'
+import type { ResolvedProfile } from './config.js'
 import type { ProviderEvent, ProviderResult, AIProvider } from '../ai-providers/types.js'
 
 export type {
@@ -81,92 +82,45 @@ export class StreamableResult implements PromiseLike<ProviderResult>, AsyncItera
 // ==================== Types ====================
 
 export interface AskOptions {
-  /**
-   * Preamble text describing the conversation context.
-   * Claude Code: injected inside the `<chat_history>` text block.
-   * Vercel AI SDK: not used (native ModelMessage[] carries the history directly).
-   */
+  /** Preamble text describing the conversation context. */
   historyPreamble?: string
-  /**
-   * System prompt override for this call.
-   * Claude Code: passed as `--system-prompt` to the CLI.
-   * Vercel AI SDK: replaces the agent's `instructions` for this call (triggers agent re-creation if changed).
-   */
+  /** System prompt override for this call. */
   systemPrompt?: string
-  /**
-   * Max text history entries to include in context.
-   * Claude Code: limits entries in the `<chat_history>` block. Default: 50.
-   * Vercel AI SDK: not used (compaction via `compactIfNeeded` controls context size).
-   */
+  /** Max text history entries to include in context (text providers only). */
   maxHistoryEntries?: number
-  /**
-   * Tool names to disable for this call, in addition to the global disabled list.
-   * Claude Code: merged into `disallowedTools` CLI option.
-   * Vercel AI SDK: filtered out from the tool map before the agent is created.
-   */
+  /** Tool names to disable for this call. */
   disabledTools?: string[]
-  /**
-   * AI provider to use for this call, overriding the global ai-provider-manager.json config.
-   * Falls back to global config if not specified.
-   */
-  provider?: 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk' | 'codex'
-  /**
-   * Vercel AI SDK model override — per-request provider/model/baseUrl/apiKey.
-   * Only used when the active backend is 'vercel-ai-sdk'.
-   */
-  vercelAiSdk?: {
-    provider: string
-    model: string
-    baseUrl?: string
-    apiKey?: string
-  }
-  /**
-   * Agent SDK model override — per-request model/apiKey/baseUrl.
-   * Only used when the active backend is 'agent-sdk'.
-   */
-  agentSdk?: {
-    model?: string
-    apiKey?: string
-    baseUrl?: string
-  }
-  /**
-   * Codex model override — per-request model/baseUrl.
-   * Only used when the active backend is 'codex'.
-   */
-  codex?: {
-    model?: string
-    baseUrl?: string
-    apiKey?: string
-    loginMethod?: 'api-key' | 'codex-oauth'
-  }
+  /** Profile slug override. Falls back to global activeProfile if omitted. */
+  profileSlug?: string
 }
 
 // ==================== GenerateRouter ====================
 
-/** Reads runtime AI config and resolves to the correct AIProvider. */
+/** Resolves profile → AIProvider instance + resolved config. */
 export class GenerateRouter {
+  private providers: Record<string, AIProvider>
+
   constructor(
-    private vercel: AIProvider,
-    private agentSdk: AIProvider | null = null,
-    private codex: AIProvider | null = null,
-  ) {}
-
-  /** Resolve the active provider, optionally overridden per-request. */
-  async resolve(override?: string): Promise<AIProvider> {
-    if (override === 'codex' && this.codex) return this.codex
-    // 'claude-code' is a legacy alias for 'agent-sdk'
-    if ((override === 'agent-sdk' || override === 'claude-code') && this.agentSdk) return this.agentSdk
-    if (override === 'vercel-ai-sdk') return this.vercel
-
-    const config = await readAIProviderConfig()
-    if (config.backend === 'codex' && this.codex) return this.codex
-    if ((config.backend === 'agent-sdk' || config.backend === 'claude-code') && this.agentSdk) return this.agentSdk
-    return this.vercel
+    vercel: AIProvider,
+    agentSdk: AIProvider | null = null,
+    codex: AIProvider | null = null,
+  ) {
+    this.providers = { 'vercel-ai-sdk': vercel }
+    if (agentSdk) this.providers['agent-sdk'] = agentSdk
+    if (codex) this.providers['codex'] = codex
+  }
+
+  /** Resolve profile and pick the matching provider. */
+  async resolve(profileSlug?: string): Promise<{ provider: AIProvider; profile: ResolvedProfile }> {
+    const profile = await resolveProfile(profileSlug)
+    const provider = this.providers[profile.backend]
+    if (!provider) throw new Error(`No provider registered for backend: ${profile.backend}`)
+    return { provider, profile }
   }
 
-  /** Stateless ask — delegates to the resolved provider. */
+  /** Stateless ask — delegates to the active profile's provider. */
   async ask(prompt: string): Promise<ProviderResult> {
-    const provider = await this.resolve()
+    const { provider } = await this.resolve()
     return provider.ask(prompt)
   }
 }
diff --git a/src/core/config.spec.ts b/src/core/config.spec.ts
index 1b06d97e..798ac14f 100644
--- a/src/core/config.spec.ts
+++ b/src/core/config.spec.ts
@@ -18,8 +18,7 @@ vi.mock('fs/promises', () => ({
 import { readFile, writeFile, mkdir } from 'fs/promises'
 import {
   readAIProviderConfig,
-  readAIBackend,
-  writeAIBackend,
+  setActiveProfile,
   readToolsConfig,
   readAgentConfig,
   readMarketDataConfig,
@@ -27,6 +26,7 @@ import {
   readAccountsConfig,
   writeAccountsConfig,
   aiProviderSchema,
+  profileSchema,
 } from './config.js'
 
 const mockReadFile = vi.mocked(readFile)
@@ -62,74 +62,55 @@ describe('readAIProviderConfig', () => {
   it('returns schema defaults when file is missing', async () => {
     fileNotFound()
     const cfg = await readAIProviderConfig()
-    expect(cfg.backend).toBe('claude-code')
-    expect(cfg.provider).toBe('anthropic')
-    expect(cfg.model).toBe('claude-sonnet-4-6')
+    expect(cfg.activeProfile).toBe('default')
+    expect(cfg.profiles.default).toBeDefined()
+    expect(cfg.profiles.default.backend).toBe('agent-sdk')
   })
 
-  it('parses valid file content', async () => {
-    fileReturns({ backend: 'vercel-ai-sdk', provider: 'openai', model: 'gpt-4o' })
+  it('parses valid profile-based content', async () => {
+    fileReturns({
+      apiKeys: { openai: 'sk-test' },
+      profiles: { main: { backend: 'codex', label: 'GPT', model: 'gpt-5.4', loginMethod: 'codex-oauth' } },
+      activeProfile: 'main',
+    })
     const cfg = await readAIProviderConfig()
-    expect(cfg.backend).toBe('vercel-ai-sdk')
-    expect(cfg.provider).toBe('openai')
-    expect(cfg.model).toBe('gpt-4o')
+    expect(cfg.activeProfile).toBe('main')
+    expect(cfg.profiles.main.backend).toBe('codex')
+    expect(cfg.profiles.main.model).toBe('gpt-5.4')
   })
 
   it('returns defaults when file contains invalid JSON (parse error)', async () => {
     fileReadError('Unexpected token')
     const cfg = await readAIProviderConfig()
-    expect(cfg.backend).toBe('claude-code')
-  })
-
-  it('fills in missing fields with schema defaults', async () => {
-    fileReturns({ backend: 'agent-sdk' })
-    const cfg = await readAIProviderConfig()
-    expect(cfg.backend).toBe('agent-sdk')
-    expect(cfg.provider).toBe('anthropic')   // default
-    expect(cfg.model).toBe('claude-sonnet-4-6') // default
+    expect(cfg.activeProfile).toBe('default')
   })
 })
 
-// ==================== readAIBackend ====================
-
-describe('readAIBackend', () => {
-  it('returns claude-code backend by default', async () => {
-    fileNotFound()
-    const { backend } = await readAIBackend()
-    expect(backend).toBe('claude-code')
-  })
-
-  it('returns the backend stored in file', async () => {
-    fileReturns({ backend: 'vercel-ai-sdk' })
-    const { backend } = await readAIBackend()
-    expect(backend).toBe('vercel-ai-sdk')
-  })
-})
-
-// ==================== writeAIBackend ====================
-
-describe('writeAIBackend', () => {
-  it('reads current config and overwrites only the backend field', async () => {
-    // First read: return existing config with custom model
-    fileReturns({ backend: 'claude-code', provider: 'anthropic', model: 'my-custom-model' })
+// ==================== setActiveProfile ====================
+
+describe('setActiveProfile', () => {
+  it('updates activeProfile and writes to disk', async () => {
+    const config = {
+      apiKeys: {},
+      profiles: {
+        a: { backend: 'agent-sdk', label: 'A', model: 'claude-sonnet-4-6', loginMethod: 'api-key' },
+        b: { backend: 'codex', label: 'B', model: 'gpt-5.4', loginMethod: 'codex-oauth' },
+      },
+      activeProfile: 'a',
+    }
+    fileReturns(config)
 
-    await writeAIBackend('vercel-ai-sdk')
+    await setActiveProfile('b')
 
-    expect(mockMkdir).toHaveBeenCalled()
     expect(mockWriteFile).toHaveBeenCalled()
-
     const written = JSON.parse((mockWriteFile.mock.calls[0][1] as string))
-    expect(written.backend).toBe('vercel-ai-sdk')
-    expect(written.model).toBe('my-custom-model') // preserved
-    expect(written.provider).toBe('anthropic')    // preserved
+    expect(written.activeProfile).toBe('b')
+    expect(written.profiles.a).toBeDefined() // preserved
   })
 
-  it('writes to ai-provider-manager.json', async () => {
-    fileReturns({ backend: 'agent-sdk' })
-    await writeAIBackend('claude-code')
-
-    const filePath = mockWriteFile.mock.calls[0][0] as string
-    expect(filePath).toMatch(/ai-provider-manager\.json$/)
+  it('throws on unknown profile slug', async () => {
+    fileReturns({ apiKeys: {}, profiles: { a: { backend: 'agent-sdk', label: 'A', model: 'x' } }, activeProfile: 'a' })
+    await expect(setActiveProfile('nonexistent')).rejects.toThrow('Unknown profile')
   })
 })
 
@@ -211,7 +192,7 @@ describe('writeConfigSection', () => {
 
   it('throws ZodError for invalid data (does not write file)', async () => {
     await expect(
-      writeConfigSection('aiProvider', { backend: 'invalid-backend-name' })
+      writeConfigSection('aiProvider', { profiles: { bad: { backend: 'invalid-backend', label: 'X' } } })
     ).rejects.toThrow()
     // writeFile should not have been called
     expect(mockWriteFile).not.toHaveBeenCalled()
@@ -270,22 +251,40 @@ describe('writeAccountsConfig', () => {
 
 // ==================== aiProviderSchema (Zod schema validation) ====================
 
-describe('aiProviderSchema', () => {
-  it('accepts valid backends', () => {
-    for (const backend of ['claude-code', 'vercel-ai-sdk', 'agent-sdk'] as const) {
-      expect(() => aiProviderSchema.parse({ backend })).not.toThrow()
-    }
+describe('aiProviderSchema (profile-based)', () => {
+  it('uses defaults for empty object', () => {
+    const result = aiProviderSchema.parse({})
+    expect(result.activeProfile).toBe('default')
+    expect(result.profiles.default).toBeDefined()
+    expect(result.apiKeys).toEqual({})
   })
 
-  it('rejects unknown backend', () => {
-    expect(() => aiProviderSchema.parse({ backend: 'unknown-backend' })).toThrow()
+  it('accepts valid profile-based config', () => {
+    expect(() => aiProviderSchema.parse({
+      profiles: { test: { backend: 'codex', label: 'Test', model: 'gpt-5.4', loginMethod: 'codex-oauth' } },
+      activeProfile: 'test',
+    })).not.toThrow()
   })
+})
 
-  it('uses defaults for missing fields', () => {
-    const result = aiProviderSchema.parse({})
-    expect(result.backend).toBe('claude-code')
-    expect(result.provider).toBe('anthropic')
-    expect(result.model).toBe('claude-sonnet-4-6')
-    expect(result.apiKeys).toEqual({})
+describe('profileSchema', () => {
+  it('validates agent-sdk profile', () => {
+    const result = profileSchema.parse({ backend: 'agent-sdk', label: 'Claude', model: 'claude-opus-4-6', loginMethod: 'claudeai' })
+    expect(result.backend).toBe('agent-sdk')
+  })
+
+  it('validates codex profile', () => {
+    const result = profileSchema.parse({ backend: 'codex', label: 'GPT', model: 'gpt-5.4' })
+    expect(result.backend).toBe('codex')
+    if (result.backend === 'codex') expect(result.loginMethod).toBe('codex-oauth') // default
+  })
+
+  it('validates vercel profile', () => {
+    const result = profileSchema.parse({ backend: 'vercel-ai-sdk', label: 'Gemini', provider: 'google', model: 'gemini-2.5-flash' })
+    expect(result.backend).toBe('vercel-ai-sdk')
+  })
+
+  it('rejects unknown backend', () => {
+    expect(() => profileSchema.parse({ backend: 'unknown', label: 'X', model: 'y' })).toThrow()
   })
 })
diff --git a/src/core/config.ts b/src/core/config.ts
index 46508c7d..9dbb477a 100644
--- a/src/core/config.ts
+++ b/src/core/config.ts
@@ -13,15 +13,17 @@ const engineSchema = z.object({
   port: z.number().int().positive().default(3000),
 })
 
-const loginMethodSchema = z.enum(['api-key', 'claudeai', 'codex-oauth'])
+// ==================== AI Provider: Legacy Schema (kept for migration) ====================
 
-export const aiProviderSchema = z.object({
+const legacyLoginMethodSchema = z.enum(['api-key', 'claudeai', 'codex-oauth'])
+
+/** @deprecated Legacy flat schema — used only for migration detection. */
+export const aiProviderLegacySchema = z.object({
   backend: z.enum(['claude-code', 'vercel-ai-sdk', 'agent-sdk', 'codex']).default('claude-code'),
   provider: z.string().default('anthropic'),
   model: z.string().default('claude-sonnet-4-6'),
   baseUrl: z.string().min(1).optional(),
-  /** Authentication method for Agent SDK: api-key (default), oauth (Console), claudeai (Pro/Max). */
-  loginMethod: loginMethodSchema.default('api-key'),
+  loginMethod: legacyLoginMethodSchema.default('api-key'),
   apiKeys: z.object({
     anthropic: z.string().optional(),
     openai: z.string().optional(),
@@ -29,6 +31,62 @@ export const aiProviderSchema = z.object({
   }).default({}),
 })
 
+// ==================== AI Provider: Profile-based Schema ====================
+
+export type AIBackend = 'agent-sdk' | 'codex' | 'vercel-ai-sdk'
+
+const apiKeysSchema = z.object({
+  anthropic: z.string().optional(),
+  openai: z.string().optional(),
+  google: z.string().optional(),
+})
+
+const baseProfileFields = {
+  label: z.string().min(1),
+  baseUrl: z.string().optional(),
+  apiKey: z.string().optional(),
+}
+
+export const agentSdkProfileSchema = z.object({
+  ...baseProfileFields,
+  backend: z.literal('agent-sdk'),
+  model: z.string().default('claude-sonnet-4-6'),
+  loginMethod: z.enum(['api-key', 'claudeai']).default('api-key'),
+})
+
+export const codexProfileSchema = z.object({
+  ...baseProfileFields,
+  backend: z.literal('codex'),
+  model: z.string().default('gpt-5.4'),
+  loginMethod: z.enum(['api-key', 'codex-oauth']).default('codex-oauth'),
+})
+
+export const vercelProfileSchema = z.object({
+  ...baseProfileFields,
+  backend: z.literal('vercel-ai-sdk'),
+  provider: z.string().default('anthropic'),
+  model: z.string().default('claude-sonnet-4-6'),
+})
+
+export const profileSchema = z.discriminatedUnion('backend', [
+  agentSdkProfileSchema, codexProfileSchema, vercelProfileSchema,
+])
+
+export type Profile = z.infer<typeof profileSchema>
+
+export const aiProviderSchema = z.object({
+  apiKeys: apiKeysSchema.default({}),
+  profiles: z.record(
+    z.string(),
+    profileSchema,
+  ).default({
+    default: { backend: 'agent-sdk', label: 'Claude Sonnet', model: 'claude-sonnet-4-6', loginMethod: 'claudeai' },
+  }),
+  activeProfile: z.string().default('default'),
+})
+
+export type AIProviderConfig = z.infer<typeof aiProviderSchema>
+
 const agentSchema = z.object({
   maxSteps: z.number().int().positive().default(20),
   evolutionMode: z.boolean().default(false),
@@ -176,43 +234,14 @@ export const toolsSchema = z.object({
   disabled: z.array(z.string()).default([]),
 })
 
-/** Vercel AI SDK model override — per-channel provider/model/key/endpoint. */
-export const vercelAiSdkOverrideSchema = z.object({
-  provider: z.string(),
-  model: z.string(),
-  baseUrl: z.string().optional(),
-  apiKey: z.string().optional(),
-})
-
-/** Agent SDK model override — per-channel model/key/endpoint. */
-export const agentSdkOverrideSchema = z.object({
-  model: z.string().optional(),
-  apiKey: z.string().optional(),
-  baseUrl: z.string().optional(),
-  loginMethod: loginMethodSchema.optional(),
-})
-
-export const codexOverrideSchema = z.object({
-  model: z.string().optional(),
-  baseUrl: z.string().optional(),
-  apiKey: z.string().optional(),
-  loginMethod: z.enum(['api-key', 'codex-oauth']).optional(),
-})
-
 export const webSubchannelSchema = z.object({
   /** URL-safe identifier. Used as session path segment: data/sessions/web/{id}.jsonl */
   id: z.string().regex(/^[a-z0-9-_]+$/, 'id must be lowercase alphanumeric with hyphens/underscores'),
   label: z.string().min(1),
   /** System prompt override for this channel. */
   systemPrompt: z.string().optional(),
-  /** AI backend override. Falls back to global config if omitted. */
-  provider: z.enum(['claude-code', 'vercel-ai-sdk', 'agent-sdk', 'codex']).optional(),
-  /** Vercel AI SDK model override. Only used when provider is 'vercel-ai-sdk'. */
-  vercelAiSdk: vercelAiSdkOverrideSchema.optional(),
-  /** Agent SDK model override. Only used when provider is 'agent-sdk'. */
-  agentSdk: agentSdkOverrideSchema.optional(),
-  /** Codex model override. Only used when provider is 'codex'. */
-  codex: codexOverrideSchema.optional(),
+  /** AI provider profile slug. Falls back to global activeProfile if omitted. */
+  profile: z.string().optional(),
   /** Tool names to disable in addition to the global disabled list. */
   disabledTools: z.array(z.string()).optional(),
 })
@@ -292,18 +321,99 @@ export async function loadConfig(): Promise<Config> {
   const raws = await Promise.all(files.map((f) => loadJsonFile(f)))
 
   // TODO: remove all migration blocks before v1.0 — no stable release yet, breaking changes are fine
-  // ---------- Migration: consolidate old ai-provider + model + api-keys → ai-provider ----------
+  // ---------- Migration: flat ai-provider config → profile-based ----------
   const aiProviderRaw = raws[6] as Record<string, unknown> | undefined
-  if (aiProviderRaw && !('backend' in aiProviderRaw)) {
-    // Old format detected — merge model.json + api-keys.json into ai-provider-manager.json
+  if (aiProviderRaw && 'backend' in aiProviderRaw && !('profiles' in aiProviderRaw)) {
+    // Legacy flat format detected — convert to profile-based
+
+    // Step 1: handle very old format (model.json + api-keys.json)
+    if (!('model' in aiProviderRaw)) {
+      const oldModel = await loadJsonFile('model.json') as Record<string, unknown> | undefined
+      const oldKeys = await loadJsonFile('api-keys.json') as Record<string, unknown> | undefined
+      if (oldModel) Object.assign(aiProviderRaw, { provider: oldModel.provider, model: oldModel.model, ...(oldModel.baseUrl ? { baseUrl: oldModel.baseUrl } : {}) })
+      if (oldKeys) aiProviderRaw.apiKeys = oldKeys
+      await removeJsonFile('model.json')
+      await removeJsonFile('api-keys.json')
+    }
+
+    // Step 2: handle claude-code → agent-sdk alias
+    if (aiProviderRaw.backend === 'claude-code') {
+      aiProviderRaw.backend = 'agent-sdk'
+      aiProviderRaw.loginMethod = aiProviderRaw.loginMethod ?? 'claudeai'
+    }
+
+    // Step 3: build default profile from flat config
+    const legacy = aiProviderLegacySchema.parse(aiProviderRaw)
+    const defaultProfile: Record<string, unknown> = { label: 'Default' }
+    if (legacy.backend === 'agent-sdk') {
+      defaultProfile.backend = 'agent-sdk'
+      defaultProfile.model = legacy.model
+      defaultProfile.loginMethod = legacy.loginMethod === 'codex-oauth' ? 'api-key' : legacy.loginMethod
+    } else if (legacy.backend === 'codex') {
+      defaultProfile.backend = 'codex'
+      defaultProfile.model = legacy.model
+      defaultProfile.loginMethod = legacy.loginMethod === 'claudeai' ? 'codex-oauth' : legacy.loginMethod
+    } else {
+      defaultProfile.backend = 'vercel-ai-sdk'
+      defaultProfile.provider = legacy.provider
+      defaultProfile.model = legacy.model
+    }
+    if (legacy.baseUrl) defaultProfile.baseUrl = legacy.baseUrl
+
+    // Step 4: migrate subchannel inline overrides → named profiles
+    const oldSubchannels = await loadJsonFile('web-subchannels.json') as Array<Record<string, unknown>> | undefined
+    const profiles: Record<string, unknown> = { default: defaultProfile }
+    const newSubchannels: Array<Record<string, unknown>> = []
+
+    if (oldSubchannels) {
+      for (const ch of oldSubchannels) {
+        const sub: Record<string, unknown> = { id: ch.id, label: ch.label }
+        if (ch.systemPrompt) sub.systemPrompt = ch.systemPrompt
+        if (ch.disabledTools) sub.disabledTools = ch.disabledTools
+
+        const provider = ch.provider as string | undefined
+        const override = provider === 'vercel-ai-sdk' ? ch.vercelAiSdk
+          : provider === 'agent-sdk' ? ch.agentSdk
+          : provider === 'codex' ? ch.codex
+          : undefined
+
+        if (provider && override) {
+          const slug = `${ch.id}-${provider}`
+          profiles[slug] = { backend: provider, label: `${ch.label}`, ...(override as object) }
+          sub.profile = slug
+        } else if (provider) {
+          // Provider set but no override — create a profile with just the backend
+          const slug = `${ch.id}-${provider}`
+          profiles[slug] = { ...defaultProfile, backend: provider, label: `${ch.label}` }
+          sub.profile = slug
+        }
+
+        newSubchannels.push(sub)
+      }
+      await writeFile(resolve(CONFIG_DIR, 'web-subchannels.json'), JSON.stringify(newSubchannels, null, 2) + '\n')
+    }
+
+    // Step 5: write new format
+    const migrated = { apiKeys: legacy.apiKeys, profiles, activeProfile: 'default' }
+    raws[6] = migrated
+    await mkdir(CONFIG_DIR, { recursive: true })
+    await writeFile(resolve(CONFIG_DIR, 'ai-provider-manager.json'), JSON.stringify(migrated, null, 2) + '\n')
+  } else if (aiProviderRaw && !('backend' in aiProviderRaw) && !('profiles' in aiProviderRaw)) {
+    // Very old format (no backend, no profiles) — handle model.json merge first
     const oldModel = await loadJsonFile('model.json') as Record<string, unknown> | undefined
     const oldKeys = await loadJsonFile('api-keys.json') as Record<string, unknown> | undefined
     const migrated = {
-      backend: aiProviderRaw.provider ?? 'claude-code',
-      provider: oldModel?.provider ?? 'anthropic',
-      model: oldModel?.model ?? 'claude-sonnet-4-6',
-      ...(oldModel?.baseUrl ? { baseUrl: oldModel.baseUrl } : {}),
       apiKeys: oldKeys ?? {},
+      profiles: {
+        default: {
+          backend: 'agent-sdk',
+          label: 'Default',
+          model: (oldModel?.model as string) ?? 'claude-sonnet-4-6',
+          loginMethod: 'claudeai',
+          provider: (oldModel?.provider as string) ?? 'anthropic',
+        },
+      },
+      activeProfile: 'default',
     }
     raws[6] = migrated
     await mkdir(CONFIG_DIR, { recursive: true })
@@ -312,14 +422,6 @@ export async function loadConfig(): Promise<Config> {
     await removeJsonFile('api-keys.json')
   }
 
-  // ---------- Migration: claude-code backend → agent-sdk + claudeai ----------
-  if (aiProviderRaw && (aiProviderRaw as Record<string, unknown>).backend === 'claude-code') {
-    const patched = { ...(aiProviderRaw as Record<string, unknown>), backend: 'agent-sdk', loginMethod: 'claudeai' }
-    raws[6] = patched
-    await mkdir(CONFIG_DIR, { recursive: true })
-    await writeFile(resolve(CONFIG_DIR, 'ai-provider-manager.json'), JSON.stringify(patched, null, 2) + '\n')
-  }
-
   // ---------- Migration: consolidate old telegram.json + engine port fields ----------
   const connectorsRaw = raws[9] as Record<string, unknown> | undefined
   if (connectorsRaw === undefined) {
@@ -452,24 +554,74 @@ export async function readConnectorsConfig() {
   }
 }
 
-// ==================== AI Backend Helpers ====================
+// ==================== Profile Helpers ====================
+
+/** Resolved profile with apiKey filled from global keys. */
+export interface ResolvedProfile {
+  backend: AIBackend
+  label: string
+  model: string
+  apiKey?: string
+  baseUrl?: string
+  loginMethod?: string
+  provider?: string
+}
 
-export type AIBackend = 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk' | 'codex'
+/** Resolve a profile by slug, filling in global apiKey fallback. */
+export async function resolveProfile(slug?: string): Promise<ResolvedProfile> {
+  const config = await readAIProviderConfig()
+  const key = slug ?? config.activeProfile
+  const profile = config.profiles[key]
+  if (!profile) throw new Error(`Unknown AI provider profile: "${key}"`)
+  const vendor = profile.backend === 'codex' ? 'openai'
+    : profile.backend === 'agent-sdk' ? 'anthropic'
+    : (profile as { provider?: string }).provider ?? 'anthropic'
+  return {
+    ...profile,
+    apiKey: profile.apiKey ?? (config.apiKeys as Record<string, string | undefined>)[vendor],
+  }
+}
 
-/** Read the current AI backend from ai-provider-manager.json. */
-export async function readAIBackend(): Promise<{ backend: AIBackend }> {
+/** Get the active profile slug. */
+export async function getActiveProfileSlug(): Promise<string> {
   const config = await readAIProviderConfig()
-  return { backend: config.backend }
+  return config.activeProfile
 }
 
-/** Switch the AI backend in ai-provider-manager.json (preserves other fields). */
-export async function writeAIBackend(backend: AIBackend): Promise<void> {
-  const current = await readAIProviderConfig()
-  const updated = { ...current, backend }
+/** Set the active profile. */
+export async function setActiveProfile(slug: string): Promise<void> {
+  const config = await readAIProviderConfig()
+  if (!config.profiles[slug]) throw new Error(`Unknown profile: "${slug}"`)
+  const updated = { ...config, activeProfile: slug }
   await mkdir(CONFIG_DIR, { recursive: true })
   await writeFile(resolve(CONFIG_DIR, 'ai-provider-manager.json'), JSON.stringify(updated, null, 2) + '\n')
 }
 
+/** Write a single profile (create or update). */
+export async function writeProfile(slug: string, profile: Profile): Promise<void> {
+  const config = await readAIProviderConfig()
+  config.profiles[slug] = profile
+  await mkdir(CONFIG_DIR, { recursive: true })
+  await writeFile(resolve(CONFIG_DIR, 'ai-provider-manager.json'), JSON.stringify(config, null, 2) + '\n')
+}
+
+/** Delete a profile. Cannot delete the active profile. */
+export async function deleteProfile(slug: string): Promise<void> {
+  const config = await readAIProviderConfig()
+  if (config.activeProfile === slug) throw new Error('Cannot delete the active profile')
+  delete config.profiles[slug]
+  await mkdir(CONFIG_DIR, { recursive: true })
+  await writeFile(resolve(CONFIG_DIR, 'ai-provider-manager.json'), JSON.stringify(config, null, 2) + '\n')
+}
+
+/** Update global API keys. */
+export async function writeApiKeys(keys: { anthropic?: string; openai?: string; google?: string }): Promise<void> {
+  const config = await readAIProviderConfig()
+  config.apiKeys = { ...config.apiKeys, ...keys }
+  await mkdir(CONFIG_DIR, { recursive: true })
+  await writeFile(resolve(CONFIG_DIR, 'ai-provider-manager.json'), JSON.stringify(config, null, 2) + '\n')
+}
+
 // ==================== Writer ====================
 
 export type ConfigSection = keyof Config

From e9565ce8df4fb25599f331f2f30851d8f68258e0 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 20:39:28 +0800
Subject: [PATCH 12/15] feat: rewrite frontend for profile-based AI config

- AIProviderPage: profile manager with create/edit/delete, active profile
  selection, per-backend form fields, global API keys section
- ChannelConfigModal: replace 3 inline override forms (~150 lines) with
  a single profile dropdown selector
- types.ts: Profile + AIBackend types, simplified WebChannel (profile slug
  instead of inline overrides), new AIProviderConfig structure
- config.ts: profile CRUD API methods (get/create/update/delete/setActive),
  API keys management, remove old setBackend()
- channels.ts: simplified ChannelListItem with profile reference

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 ui/src/api/channels.ts                   |  19 +-
 ui/src/api/config.ts                     |  82 +-
 ui/src/api/index.ts                      |   4 +-
 ui/src/api/types.ts                      |  39 +-
 ui/src/components/ChannelConfigModal.tsx | 336 +-------
 ui/src/pages/AIProviderPage.tsx          | 983 ++++++++---------------
 6 files changed, 478 insertions(+), 985 deletions(-)

diff --git a/ui/src/api/channels.ts b/ui/src/api/channels.ts
index 3479ad4c..ea91da3b 100644
--- a/ui/src/api/channels.ts
+++ b/ui/src/api/channels.ts
@@ -1,14 +1,11 @@
 import { headers } from './client'
-import type { WebChannel, VercelAiSdkOverride, AgentSdkOverride, CodexOverride } from './types'
+import type { WebChannel } from './types'
 
 export interface ChannelListItem {
   id: string
   label: string
   systemPrompt?: string
-  provider?: 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk' | 'codex'
-  vercelAiSdk?: VercelAiSdkOverride
-  agentSdk?: AgentSdkOverride
-  codex?: CodexOverride
+  profile?: string
   disabledTools?: string[]
 }
 
@@ -33,7 +30,7 @@ export const channelsApi = {
   },
 
   async update(id: string, data: Partial<Omit<WebChannel, 'id'>>): Promise<{ channel: ChannelListItem }> {
-    const res = await fetch(`/api/channels/${encodeURIComponent(id)}`, {
+    const res = await fetch(`/api/channels/${id}`, {
       method: 'PUT',
       headers,
       body: JSON.stringify(data),
@@ -46,13 +43,7 @@ export const channelsApi = {
   },
 
   async remove(id: string): Promise<void> {
-    const res = await fetch(`/api/channels/${encodeURIComponent(id)}`, {
-      method: 'DELETE',
-      headers,
-    })
-    if (!res.ok) {
-      const err = await res.json().catch(() => ({ error: res.statusText }))
-      throw new Error(err.error || res.statusText)
-    }
+    const res = await fetch(`/api/channels/${id}`, { method: 'DELETE' })
+    if (!res.ok) throw new Error('Failed to delete channel')
   },
 }
diff --git a/ui/src/api/config.ts b/ui/src/api/config.ts
index cf1c2285..08ff00c7 100644
--- a/ui/src/api/config.ts
+++ b/ui/src/api/config.ts
@@ -1,5 +1,5 @@
 import { headers } from './client'
-import type { AppConfig } from './types'
+import type { AppConfig, Profile } from './types'
 
 export const configApi = {
   async load(): Promise<AppConfig> {
@@ -8,15 +8,6 @@ export const configApi = {
     return res.json()
   },
 
-  async setBackend(backend: string): Promise<void> {
-    const res = await fetch('/api/config/ai-provider', {
-      method: 'PUT',
-      headers,
-      body: JSON.stringify({ backend }),
-    })
-    if (!res.ok) throw new Error('Failed to switch backend')
-  },
-
   async updateSection(section: string, data: unknown): Promise<unknown> {
     const res = await fetch(`/api/config/${section}`, {
       method: 'PUT',
@@ -29,4 +20,75 @@ export const configApi = {
     }
     return res.json()
   },
+
+  // ==================== Profile CRUD ====================
+
+  async getProfiles(): Promise<{ profiles: Record<string, Profile>; activeProfile: string }> {
+    const res = await fetch('/api/config/profiles')
+    if (!res.ok) throw new Error('Failed to load profiles')
+    return res.json()
+  },
+
+  async createProfile(slug: string, profile: Profile): Promise<{ slug: string; profile: Profile }> {
+    const res = await fetch('/api/config/profiles', {
+      method: 'POST',
+      headers,
+      body: JSON.stringify({ slug, profile }),
+    })
+    if (!res.ok) {
+      const err = await res.json().catch(() => ({ error: 'Failed to create profile' }))
+      throw new Error(err.error || 'Failed to create profile')
+    }
+    return res.json()
+  },
+
+  async updateProfile(slug: string, profile: Profile): Promise<{ slug: string; profile: Profile }> {
+    const res = await fetch(`/api/config/profiles/${slug}`, {
+      method: 'PUT',
+      headers,
+      body: JSON.stringify(profile),
+    })
+    if (!res.ok) {
+      const err = await res.json().catch(() => ({ error: 'Failed to update profile' }))
+      throw new Error(err.error || 'Failed to update profile')
+    }
+    return res.json()
+  },
+
+  async deleteProfile(slug: string): Promise<void> {
+    const res = await fetch(`/api/config/profiles/${slug}`, { method: 'DELETE' })
+    if (!res.ok) {
+      const err = await res.json().catch(() => ({ error: 'Failed to delete profile' }))
+      throw new Error(err.error || 'Failed to delete profile')
+    }
+  },
+
+  async setActiveProfile(slug: string): Promise<void> {
+    const res = await fetch('/api/config/active-profile', {
+      method: 'PUT',
+      headers,
+      body: JSON.stringify({ slug }),
+    })
+    if (!res.ok) {
+      const err = await res.json().catch(() => ({ error: 'Failed to set active profile' }))
+      throw new Error(err.error || 'Failed to set active profile')
+    }
+  },
+
+  // ==================== API Keys ====================
+
+  async updateApiKeys(keys: { anthropic?: string; openai?: string; google?: string }): Promise<void> {
+    const res = await fetch('/api/config/api-keys', {
+      method: 'PUT',
+      headers,
+      body: JSON.stringify(keys),
+    })
+    if (!res.ok) throw new Error('Failed to update API keys')
+  },
+
+  async getApiKeysStatus(): Promise<{ anthropic: boolean; openai: boolean; google: boolean }> {
+    const res = await fetch('/api/config/api-keys/status')
+    if (!res.ok) throw new Error('Failed to load API key status')
+    return res.json()
+  },
 }
diff --git a/ui/src/api/index.ts b/ui/src/api/index.ts
index 94fed442..dbe1f73f 100644
--- a/ui/src/api/index.ts
+++ b/ui/src/api/index.ts
@@ -32,7 +32,8 @@ export const api = {
 // Re-export all types for convenience
 export type {
   WebChannel,
-  VercelAiSdkOverride,
+  Profile,
+  AIBackend,
   ChatMessage,
   ChatResponse,
   ToolCall,
@@ -53,7 +54,6 @@ export type {
   NewsCollectorConfig,
   NewsCollectorFeed,
   ToolCallRecord,
-  LoginMethod,
   UTASnapshotSummary,
   EquityCurvePoint,
 } from './types'
diff --git a/ui/src/api/types.ts b/ui/src/api/types.ts
index 03f9c946..ba3848a8 100644
--- a/ui/src/api/types.ts
+++ b/ui/src/api/types.ts
@@ -1,36 +1,24 @@
-// ==================== Channels ====================
-
-export interface VercelAiSdkOverride {
-  provider: string
-  model: string
-  baseUrl?: string
-  apiKey?: string
-}
+// ==================== AI Provider Profiles ====================
 
-export type LoginMethod = 'api-key' | 'claudeai' | 'codex-oauth'
+export type AIBackend = 'agent-sdk' | 'codex' | 'vercel-ai-sdk'
 
-export interface AgentSdkOverride {
-  model?: string
+export interface Profile {
+  backend: AIBackend
+  label: string
+  model: string
+  loginMethod?: string
+  provider?: string   // vercel-ai-sdk only
   baseUrl?: string
   apiKey?: string
-  loginMethod?: LoginMethod
 }
 
-export interface CodexOverride {
-  model?: string
-  baseUrl?: string
-  apiKey?: string
-  loginMethod?: 'api-key' | 'codex-oauth'
-}
+// ==================== Channels ====================
 
 export interface WebChannel {
   id: string
   label: string
   systemPrompt?: string
-  provider?: 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk' | 'codex'
-  vercelAiSdk?: VercelAiSdkOverride
-  agentSdk?: AgentSdkOverride
-  codex?: CodexOverride
+  profile?: string    // slug reference to a profile
   disabledTools?: string[]
 }
 
@@ -68,12 +56,9 @@ export type ChatHistoryItem =
 // ==================== Config ====================
 
 export interface AIProviderConfig {
-  backend: string
-  provider: string
-  model: string
-  baseUrl?: string
-  loginMethod?: LoginMethod
   apiKeys: { anthropic?: string; openai?: string; google?: string }
+  profiles: Record<string, Profile>
+  activeProfile: string
 }
 
 export interface AppConfig {
diff --git a/ui/src/components/ChannelConfigModal.tsx b/ui/src/components/ChannelConfigModal.tsx
index 73b8a0a0..9f2d18a6 100644
--- a/ui/src/components/ChannelConfigModal.tsx
+++ b/ui/src/components/ChannelConfigModal.tsx
@@ -1,6 +1,6 @@
 import { useState, useEffect } from 'react'
 import { api } from '../api'
-import type { LoginMethod } from '../api/types'
+import type { Profile } from '../api/types'
 import type { ChannelListItem } from '../api/channels'
 import type { ToolInfo } from '../api/tools'
 
@@ -13,76 +13,26 @@ interface ChannelConfigModalProps {
 export function ChannelConfigModal({ channel, onClose, onSaved }: ChannelConfigModalProps) {
   const [label, setLabel] = useState(channel.label)
   const [systemPrompt, setSystemPrompt] = useState(channel.systemPrompt ?? '')
-  const [provider, setProvider] = useState(channel.provider ?? '')
+  const [profile, setProfile] = useState(channel.profile ?? '')
   const [disabledTools, setDisabledTools] = useState<Set<string>>(new Set(channel.disabledTools ?? []))
   const [tools, setTools] = useState<ToolInfo[]>([])
+  const [profiles, setProfiles] = useState<Record<string, Profile>>({})
   const [saving, setSaving] = useState(false)
   const [error, setError] = useState('')
 
-  // Vercel AI SDK override state
-  const [vModelProvider, setVModelProvider] = useState(channel.vercelAiSdk?.provider ?? '')
-  const [vModel, setVModel] = useState(channel.vercelAiSdk?.model ?? '')
-  const [vBaseUrl, setVBaseUrl] = useState(channel.vercelAiSdk?.baseUrl ?? '')
-  const [vApiKey, setVApiKey] = useState(channel.vercelAiSdk?.apiKey ?? '')
-
-  // Agent SDK override state
-  const [aModel, setAModel] = useState(channel.agentSdk?.model ?? '')
-  const [aBaseUrl, setABaseUrl] = useState(channel.agentSdk?.baseUrl ?? '')
-  const [aApiKey, setAApiKey] = useState(channel.agentSdk?.apiKey ?? '')
-  const [aLoginMethod, setALoginMethod] = useState<LoginMethod | ''>(channel.agentSdk?.loginMethod ?? '')
-
-  // Codex override state
-  const [cModel, setCModel] = useState(channel.codex?.model ?? '')
-  const [cBaseUrl, setCBaseUrl] = useState(channel.codex?.baseUrl ?? '')
-  const [cApiKey, setCApiKey] = useState(channel.codex?.apiKey ?? '')
-  const [cLoginMethod, setCLoginMethod] = useState<'api-key' | 'codex-oauth' | ''>(channel.codex?.loginMethod ?? '')
-
-  const showVercelConfig = provider === 'vercel-ai-sdk'
-  const showAgentSdkConfig = provider === 'agent-sdk'
-  const showCodexConfig = provider === 'codex'
-
   useEffect(() => {
     api.tools.load().then(({ inventory }) => setTools(inventory)).catch(() => {})
+    api.config.getProfiles().then(({ profiles: p }) => setProfiles(p)).catch(() => {})
   }, [])
 
   const handleSave = async () => {
     setSaving(true)
     setError('')
     try {
-      const vercelAiSdk = showVercelConfig && vModelProvider && vModel
-        ? {
-            provider: vModelProvider,
-            model: vModel,
-            ...(vBaseUrl ? { baseUrl: vBaseUrl } : {}),
-            ...(vApiKey ? { apiKey: vApiKey } : {}),
-          }
-        : undefined
-
-      const agentSdk = showAgentSdkConfig && (aModel || aLoginMethod)
-        ? {
-            ...(aModel ? { model: aModel } : {}),
-            ...(aBaseUrl ? { baseUrl: aBaseUrl } : {}),
-            ...(aApiKey ? { apiKey: aApiKey } : {}),
-            ...(aLoginMethod ? { loginMethod: aLoginMethod } : {}),
-          }
-        : undefined
-
-      const codex = showCodexConfig && (cModel || cLoginMethod)
-        ? {
-            ...(cModel ? { model: cModel } : {}),
-            ...(cBaseUrl ? { baseUrl: cBaseUrl } : {}),
-            ...(cApiKey ? { apiKey: cApiKey } : {}),
-            ...(cLoginMethod ? { loginMethod: cLoginMethod as 'api-key' | 'codex-oauth' } : {}),
-          }
-        : undefined
-
       const { channel: updated } = await api.channels.update(channel.id, {
         label: label.trim() || channel.label,
         systemPrompt: systemPrompt.trim() || undefined,
-        provider: (provider as 'claude-code' | 'vercel-ai-sdk' | 'agent-sdk' | 'codex') || undefined,
-        vercelAiSdk: vercelAiSdk ?? (null as unknown as undefined),
-        agentSdk: agentSdk ?? (null as unknown as undefined),
-        codex: codex ?? (null as unknown as undefined),
+        profile: profile || undefined,
         disabledTools: disabledTools.size > 0 ? [...disabledTools] : undefined,
       })
       onSaved(updated)
@@ -102,43 +52,28 @@ export function ChannelConfigModal({ channel, onClose, onSaved }: ChannelConfigM
     })
   }
 
-  // Group tools by group name
-  const toolGroups = tools.reduce<Record<string, ToolInfo[]>>((acc, t) => {
-    ;(acc[t.group] ??= []).push(t)
-    return acc
-  }, {})
-
-  const inputClass = 'w-full text-sm px-3 py-2 rounded-lg border border-border bg-bg-secondary text-text placeholder:text-text-muted focus:outline-none focus:border-accent'
+  const inputClass = 'w-full px-3 py-2 text-[13px] rounded-lg border border-border bg-bg-secondary text-text focus:outline-none focus:ring-1 focus:ring-accent'
 
   return (
-    <div className="fixed inset-0 z-50 flex items-center justify-center bg-black/50" onClick={onClose}>
+    <div className="fixed inset-0 z-50 flex items-center justify-center bg-black/50 backdrop-blur-sm" onClick={onClose}>
       <div
         className="bg-bg border border-border rounded-xl shadow-2xl w-full max-w-lg max-h-[80vh] flex flex-col"
         onClick={(e) => e.stopPropagation()}
       >
         {/* Header */}
-        <div className="flex items-center justify-between px-5 py-4 border-b border-border">
-          <h2 className="text-base font-semibold text-text">
-            <span className="text-text-muted mr-1">#</span>
-            {channel.id}
-          </h2>
-          <button
-            onClick={onClose}
-            className="w-6 h-6 rounded flex items-center justify-center text-text-muted hover:text-text hover:bg-bg-secondary transition-colors"
-          >
-            <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="2.5">
-              <path d="M18 6L6 18M6 6l12 12" />
-            </svg>
+        <div className="flex items-center justify-between p-4 border-b border-border">
+          <h2 className="text-[15px] font-semibold text-text">Configure Channel</h2>
+          <button onClick={onClose} className="text-text-muted hover:text-text transition-colors">
+            <svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="2" strokeLinecap="round" strokeLinejoin="round"><line x1="18" y1="6" x2="6" y2="18" /><line x1="6" y1="6" x2="18" y2="18" /></svg>
           </button>
         </div>
 
         {/* Body */}
-        <div className="flex-1 overflow-y-auto px-5 py-4 space-y-4">
+        <div className="flex-1 overflow-y-auto p-4 space-y-4">
           {/* Label */}
           <div>
-            <label className="block text-xs font-medium text-text-muted mb-1">Label</label>
+            <label className="block text-xs font-medium text-text-muted mb-1">Channel Name</label>
             <input
-              type="text"
               value={label}
               onChange={(e) => setLabel(e.target.value)}
               className={inputClass}
@@ -157,222 +92,48 @@ export function ChannelConfigModal({ channel, onClose, onSaved }: ChannelConfigM
             />
           </div>
 
-          {/* AI Backend */}
+          {/* AI Profile */}
           <div>
-            <label className="block text-xs font-medium text-text-muted mb-1">AI Backend</label>
+            <label className="block text-xs font-medium text-text-muted mb-1">AI Provider Profile</label>
             <select
-              value={provider}
-              onChange={(e) => setProvider(e.target.value)}
+              value={profile}
+              onChange={(e) => setProfile(e.target.value)}
               className={inputClass}
             >
-              <option value="">Default (global)</option>
-              <option value="agent-sdk">Claude</option>
-              <option value="vercel-ai-sdk">Vercel AI SDK</option>
-              <option value="codex">Codex</option>
+              <option value="">Default (global active)</option>
+              {Object.entries(profiles).map(([slug, p]) => (
+                <option key={slug} value={slug}>{p.label} ({p.model})</option>
+              ))}
             </select>
           </div>
 
-          {/* Vercel AI SDK config — only when provider is vercel-ai-sdk */}
-          {showVercelConfig && (
-            <div className="rounded-lg border border-border/50 bg-bg-secondary/30 p-3 space-y-3">
-              <p className="text-xs font-medium text-text-muted">Vercel AI SDK Model Config</p>
-
-              <div className="grid grid-cols-2 gap-3">
-                <div>
-                  <label className="block text-xs text-text-muted/70 mb-1">LLM Provider</label>
-                  <select
-                    value={vModelProvider}
-                    onChange={(e) => setVModelProvider(e.target.value)}
-                    className={inputClass}
-                  >
-                    <option value="">Select...</option>
-                    <option value="anthropic">Anthropic</option>
-                    <option value="openai">OpenAI</option>
-                    <option value="google">Google</option>
-                  </select>
-                </div>
-                <div>
-                  <label className="block text-xs text-text-muted/70 mb-1">Model</label>
-                  <input
-                    type="text"
-                    value={vModel}
-                    onChange={(e) => setVModel(e.target.value)}
-                    placeholder="e.g. gpt-4o"
-                    className={inputClass}
-                  />
-                </div>
-              </div>
-
-              <div>
-                <label className="block text-xs text-text-muted/70 mb-1">Base URL <span className="text-text-muted/40">(optional)</span></label>
-                <input
-                  type="text"
-                  value={vBaseUrl}
-                  onChange={(e) => setVBaseUrl(e.target.value)}
-                  placeholder="https://api.openai.com/v1"
-                  className={inputClass}
-                />
-                {vModelProvider === 'anthropic' && (
-                  <p className="mt-1 text-xs text-text-muted/60">
-                    Anthropic-compatible APIs: URL must end with <code className="font-mono">/v1</code> — e.g. <code className="font-mono">https://example.com/anthropic/v1</code>
-                  </p>
-                )}
-              </div>
-
-              <div>
-                <label className="block text-xs text-text-muted/70 mb-1">API Key <span className="text-text-muted/40">(optional, overrides global)</span></label>
-                <input
-                  type="password"
-                  value={vApiKey}
-                  onChange={(e) => setVApiKey(e.target.value)}
-                  placeholder="sk-..."
-                  className={inputClass}
-                />
-              </div>
-            </div>
-          )}
-
-          {/* Agent SDK config — only when provider is agent-sdk */}
-          {showAgentSdkConfig && (
-            <div className="rounded-lg border border-border/50 bg-bg-secondary/30 p-3 space-y-3">
-              <p className="text-xs font-medium text-text-muted">Claude Override</p>
-
-              <div>
-                <label className="block text-xs text-text-muted/70 mb-1">Login Method</label>
-                <select
-                  value={aLoginMethod}
-                  onChange={(e) => setALoginMethod((e.target.value || '') as LoginMethod | '')}
-                  className={inputClass}
-                >
-                  <option value="">Default (global)</option>
-                  <option value="api-key">API Key</option>
-                  <option value="claudeai">Claude Pro/Max</option>
-                </select>
-              </div>
-
-              <div>
-                <label className="block text-xs text-text-muted/70 mb-1">Model</label>
-                <input
-                  type="text"
-                  value={aModel}
-                  onChange={(e) => setAModel(e.target.value)}
-                  placeholder="e.g. claude-opus-4-6"
-                  className={inputClass}
-                />
-              </div>
-
-              <div>
-                <label className="block text-xs text-text-muted/70 mb-1">Base URL <span className="text-text-muted/40">(optional)</span></label>
-                <input
-                  type="text"
-                  value={aBaseUrl}
-                  onChange={(e) => setABaseUrl(e.target.value)}
-                  placeholder="Leave empty for default"
-                  className={inputClass}
-                />
-              </div>
-
-              <div>
-                <label className="block text-xs text-text-muted/70 mb-1">API Key <span className="text-text-muted/40">(optional, overrides global)</span></label>
-                <input
-                  type="password"
-                  value={aApiKey}
-                  onChange={(e) => setAApiKey(e.target.value)}
-                  placeholder="sk-ant-..."
-                  className={inputClass}
-                />
-              </div>
-            </div>
-          )}
-
-          {/* Codex config — only when provider is codex */}
-          {showCodexConfig && (
-            <div className="rounded-lg border border-border/50 bg-bg-secondary/30 p-3 space-y-3">
-              <p className="text-xs font-medium text-text-muted">Codex Override</p>
-
-              <div>
-                <label className="block text-xs text-text-muted/70 mb-1">Login Method</label>
-                <select
-                  value={cLoginMethod}
-                  onChange={(e) => setCLoginMethod((e.target.value || '') as 'api-key' | 'codex-oauth' | '')}
-                  className={inputClass}
-                >
-                  <option value="">Default (global)</option>
-                  <option value="codex-oauth">ChatGPT Subscription</option>
-                  <option value="api-key">API Key</option>
-                </select>
-              </div>
-
-              <div>
-                <label className="block text-xs text-text-muted/70 mb-1">Model</label>
-                <input
-                  type="text"
-                  value={cModel}
-                  onChange={(e) => setCModel(e.target.value)}
-                  placeholder="e.g. gpt-5.4"
-                  className={inputClass}
-                />
-              </div>
-
-              <div>
-                <label className="block text-xs text-text-muted/70 mb-1">Base URL <span className="text-text-muted/40">(optional)</span></label>
-                <input
-                  type="text"
-                  value={cBaseUrl}
-                  onChange={(e) => setCBaseUrl(e.target.value)}
-                  placeholder="Leave empty for default"
-                  className={inputClass}
-                />
-              </div>
-
-              <div>
-                <label className="block text-xs text-text-muted/70 mb-1">API Key <span className="text-text-muted/40">(optional, overrides global)</span></label>
-                <input
-                  type="password"
-                  value={cApiKey}
-                  onChange={(e) => setCApiKey(e.target.value)}
-                  placeholder="sk-..."
-                  className={inputClass}
-                />
-              </div>
-            </div>
-          )}
-
           {/* Disabled Tools */}
           <div>
             <label className="block text-xs font-medium text-text-muted mb-2">
               Disabled Tools
               {disabledTools.size > 0 && (
-                <span className="ml-2 text-text-muted/60">({disabledTools.size} disabled)</span>
+                <span className="ml-1 text-text-muted/60">({disabledTools.size} disabled)</span>
               )}
             </label>
             {tools.length === 0 ? (
-              <p className="text-xs text-text-muted">Loading tools...</p>
+              <p className="text-[11px] text-text-muted">Loading tools...</p>
             ) : (
-              <div className="space-y-3 max-h-48 overflow-y-auto">
-                {Object.entries(toolGroups).map(([group, groupTools]) => (
-                  <div key={group}>
-                    <p className="text-xs font-medium text-text-muted/70 mb-1">{group}</p>
-                    <div className="flex flex-wrap gap-1.5">
-                      {groupTools.map((t) => {
-                        const isDisabled = disabledTools.has(t.name)
-                        return (
-                          <button
-                            key={t.name}
-                            onClick={() => toggleTool(t.name)}
-                            title={t.description}
-                            className={`text-xs px-2 py-1 rounded-md border transition-colors ${
-                              isDisabled
-                                ? 'border-red-400/30 bg-red-400/10 text-red-400 line-through'
-                                : 'border-border bg-bg-secondary text-text hover:border-accent/50'
-                            }`}
-                          >
-                            {t.name}
-                          </button>
-                        )
-                      })}
-                    </div>
-                  </div>
+              <div className="max-h-[200px] overflow-y-auto border border-border rounded-lg">
+                {tools.map((tool) => (
+                  <label
+                    key={tool.name}
+                    className="flex items-center gap-2 px-3 py-1.5 hover:bg-bg-tertiary cursor-pointer text-[12px]"
+                  >
+                    <input
+                      type="checkbox"
+                      checked={disabledTools.has(tool.name)}
+                      onChange={() => toggleTool(tool.name)}
+                      className="accent-accent"
+                    />
+                    <span className={disabledTools.has(tool.name) ? 'line-through text-text-muted' : 'text-text'}>
+                      {tool.name}
+                    </span>
+                  </label>
                 ))}
               </div>
             )}
@@ -380,20 +141,11 @@ export function ChannelConfigModal({ channel, onClose, onSaved }: ChannelConfigM
         </div>
 
         {/* Footer */}
-        <div className="px-5 py-3 border-t border-border flex items-center justify-between">
-          {error ? <p className="text-xs text-red-400">{error}</p> : <div />}
-          <div className="flex items-center gap-2">
-            <button
-              onClick={onClose}
-              className="text-sm px-3 py-1.5 rounded-lg text-text-muted hover:text-text transition-colors"
-            >
-              Cancel
-            </button>
-            <button
-              onClick={handleSave}
-              disabled={saving}
-              className="btn-primary"
-            >
+        <div className="flex items-center justify-between p-4 border-t border-border">
+          {error && <p className="text-[12px] text-red flex-1">{error}</p>}
+          <div className="flex gap-2 ml-auto">
+            <button onClick={onClose} className="btn-secondary">Cancel</button>
+            <button onClick={handleSave} disabled={saving} className="btn-primary">
               {saving ? 'Saving...' : 'Save'}
             </button>
           </div>
diff --git a/ui/src/pages/AIProviderPage.tsx b/ui/src/pages/AIProviderPage.tsx
index 6728466e..9f1b8d51 100644
--- a/ui/src/pages/AIProviderPage.tsx
+++ b/ui/src/pages/AIProviderPage.tsx
@@ -1,729 +1,432 @@
-import { useState, useEffect, useCallback, useRef, useMemo } from 'react'
-import { api, type AppConfig, type AIProviderConfig, type LoginMethod } from '../api'
+import { useState, useEffect, useRef } from 'react'
+import { api, type Profile, type AIBackend } from '../api'
 import { SaveIndicator } from '../components/SaveIndicator'
 import { ConfigSection, Field, inputClass } from '../components/form'
-import { useAutoSave, type SaveStatus } from '../hooks/useAutoSave'
+import type { SaveStatus } from '../hooks/useAutoSave'
 import { PageHeader } from '../components/PageHeader'
 import { PageLoading } from '../components/StateViews'
 
-const LOGIN_METHODS: { value: LoginMethod; label: string; subtitle: string; hint: string }[] = [
-  { value: 'claudeai', label: 'Claude Pro/Max', subtitle: 'Use your Claude subscription', hint: 'Requires local Claude Code login (run claude login in terminal). No API key needed.' },
-  { value: 'api-key', label: 'API Key', subtitle: 'Pay per token', hint: 'Enter your Anthropic API key below. Billed per token to your API account.' },
-]
-
-const PROVIDER_MODELS: Record<string, { label: string; value: string }[]> = {
-  anthropic: [
-    { label: 'Claude Opus 4.6', value: 'claude-opus-4-6' },
-    { label: 'Claude Sonnet 4.6', value: 'claude-sonnet-4-6' },
-    { label: 'Claude Haiku 4.5', value: 'claude-haiku-4-5' },
-  ],
-  openai: [
-    { label: 'GPT-5.2 Pro', value: 'gpt-5.2-pro' },
-    { label: 'GPT-5.2', value: 'gpt-5.2' },
-    { label: 'GPT-5 Mini', value: 'gpt-5-mini' },
-  ],
-  google: [
-    { label: 'Gemini 3.1 Pro', value: 'gemini-3.1-pro-preview' },
-    { label: 'Gemini 3 Flash', value: 'gemini-3-flash-preview' },
-    { label: 'Gemini 2.5 Pro', value: 'gemini-2.5-pro' },
-  ],
+// ==================== Constants ====================
+
+const BACKEND_INFO: Record<AIBackend, { label: string; icon: React.ReactNode }> = {
+  'agent-sdk': {
+    label: 'Claude',
+    icon: <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><path d="M12 2a4 4 0 0 1 4 4v1a1 1 0 0 1-1 1H9a1 1 0 0 1-1-1V6a4 4 0 0 1 4-4z" /><path d="M8 8v2a4 4 0 0 0 8 0V8" /><path d="M12 14v4" /><path d="M8 22h8" /><circle cx="9" cy="5.5" r="0.5" fill="currentColor" stroke="none" /><circle cx="15" cy="5.5" r="0.5" fill="currentColor" stroke="none" /></svg>,
+  },
+  'codex': {
+    label: 'OpenAI / Codex',
+    icon: <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><polyline points="16 18 22 12 16 6" /><polyline points="8 6 2 12 8 18" /><line x1="14" y1="4" x2="10" y2="20" /></svg>,
+  },
+  'vercel-ai-sdk': {
+    label: 'Vercel AI SDK',
+    icon: <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><path d="M13 10V3L4 14h7v7l9-11h-7z" /></svg>,
+  },
 }
 
-const PROVIDERS = [
-  { value: 'anthropic', label: 'Anthropic' },
-  { value: 'openai', label: 'OpenAI' },
-  { value: 'google', label: 'Google' },
-  { value: 'custom', label: 'Custom' },
-]
-
-const SDK_FORMATS = [
-  { value: 'openai', label: 'OpenAI Compatible' },
-  { value: 'anthropic', label: 'Anthropic Compatible' },
-  { value: 'google', label: 'Google Compatible' },
-]
-
-/** Detect whether saved config should show as "Custom" in the UI. */
-function detectCustomMode(provider: string, model: string): boolean {
-  const presets = PROVIDER_MODELS[provider]
-  if (!presets) return true
-  return !presets.some((p) => p.value === model)
+const NEW_PROFILE_DEFAULTS: Record<AIBackend, Omit<Profile, 'label'>> = {
+  'agent-sdk':     { backend: 'agent-sdk', model: 'claude-sonnet-4-6', loginMethod: 'claudeai' },
+  'codex':         { backend: 'codex', model: 'gpt-5.4', loginMethod: 'codex-oauth' },
+  'vercel-ai-sdk': { backend: 'vercel-ai-sdk', model: 'claude-sonnet-4-6', provider: 'anthropic' },
 }
 
-/** UI-level backend selector. */
-type UIBackend = 'agent-sdk' | 'codex' | 'vercel-ai-sdk'
-
-/** Default provider/model per UI backend — applied on every switch to avoid stale config. */
-const BACKEND_DEFAULTS: Record<UIBackend, { backend: string; provider: string; model: string }> = {
-  'agent-sdk':    { backend: 'agent-sdk',    provider: 'anthropic', model: 'claude-sonnet-4-6' },
-  'codex':        { backend: 'codex',        provider: 'openai',   model: 'gpt-5.4' },
-  'vercel-ai-sdk': { backend: 'vercel-ai-sdk', provider: 'anthropic', model: PROVIDER_MODELS.anthropic[0].value },
-}
-
-/** Derive initial UI backend from config. */
-function detectUIBackend(config: AIProviderConfig): UIBackend {
-  if (config.backend === 'codex') return 'codex'
-  // Legacy: openai via vercel-ai-sdk → migrate to codex
-  if (config.backend === 'vercel-ai-sdk' && config.provider === 'openai') return 'codex'
-  if (config.backend === 'vercel-ai-sdk') return 'vercel-ai-sdk'
-  return 'agent-sdk'
-}
-
-function BackendCard({ selected, onClick, icon, title, description }: {
-  selected: boolean
-  onClick: () => void
-  icon: React.ReactNode
-  title: string
-  description: string
-}) {
-  return (
-    <button
-      onClick={onClick}
-      className={`flex flex-col items-start gap-2 p-4 rounded-xl border-2 text-left transition-all cursor-pointer ${
-        selected
-          ? 'border-accent bg-accent-dim/50'
-          : 'border-border bg-bg hover:border-text-muted/30 hover:bg-bg-tertiary'
-      }`}
-    >
-      <div className={`${selected ? 'text-accent' : 'text-text-muted'}`}>{icon}</div>
-      <div>
-        <p className={`text-[13px] font-semibold ${selected ? 'text-accent' : 'text-text'}`}>{title}</p>
-        <p className="text-[11px] text-text-muted mt-0.5 leading-snug">{description}</p>
-      </div>
-    </button>
-  )
-}
+// ==================== Main Page ====================
 
 export function AIProviderPage() {
-  const [config, setConfig] = useState<AppConfig | null>(null)
+  const [profiles, setProfiles] = useState<Record<string, Profile> | null>(null)
+  const [activeProfile, setActiveProfile] = useState('')
+  const [apiKeys, setApiKeys] = useState<{ anthropic?: string; openai?: string; google?: string }>({})
+  const [selectedSlug, setSelectedSlug] = useState<string | null>(null)
+  const [creating, setCreating] = useState<AIBackend | null>(null)
 
   useEffect(() => {
-    api.config.load().then(setConfig).catch(() => {})
+    api.config.getProfiles().then(({ profiles: p, activeProfile: a }) => {
+      setProfiles(p)
+      setActiveProfile(a)
+      setSelectedSlug(a)
+    }).catch(() => {})
+    api.config.getApiKeysStatus().then((status) => {
+      setApiKeys({
+        ...(status.anthropic ? { anthropic: '(set)' } : {}),
+        ...(status.openai ? { openai: '(set)' } : {}),
+        ...(status.google ? { google: '(set)' } : {}),
+      })
+    }).catch(() => {})
   }, [])
 
-  const uiBackend: UIBackend = config ? detectUIBackend(config.aiProvider) : 'agent-sdk'
-
-  const handleBackendSwitch = useCallback(
-    async (target: UIBackend) => {
-      if (!config) return
-      try {
-        const defaults = BACKEND_DEFAULTS[target]
-        const updated = { ...config.aiProvider, ...defaults }
-        await api.config.updateSection('aiProvider', updated)
-        setConfig((c) => c ? { ...c, aiProvider: updated } : c)
-      } catch {
-        // Button state reflects actual saved state
-      }
-    },
-    [config],
-  )
+  const handleSetActive = async (slug: string) => {
+    try {
+      await api.config.setActiveProfile(slug)
+      setActiveProfile(slug)
+    } catch { /* keep old state */ }
+  }
+
+  const handleDelete = async (slug: string) => {
+    if (!profiles) return
+    try {
+      await api.config.deleteProfile(slug)
+      const updated = { ...profiles }
+      delete updated[slug]
+      setProfiles(updated)
+      if (selectedSlug === slug) setSelectedSlug(activeProfile)
+    } catch { /* keep old state */ }
+  }
+
+  const handleCreateStart = (backend: AIBackend) => {
+    setCreating(backend)
+    setSelectedSlug(null)
+  }
+
+  const handleCreateSave = async (slug: string, profile: Profile) => {
+    try {
+      await api.config.createProfile(slug, profile)
+      setProfiles((p) => p ? { ...p, [slug]: profile } : p)
+      setCreating(null)
+      setSelectedSlug(slug)
+    } catch { /* form handles error */ }
+  }
+
+  const handleProfileUpdate = async (slug: string, profile: Profile) => {
+    try {
+      await api.config.updateProfile(slug, profile)
+      setProfiles((p) => p ? { ...p, [slug]: profile } : p)
+    } catch { /* form handles error */ }
+  }
+
+  if (!profiles) return <div className="flex flex-col flex-1 min-h-0"><PageHeader title="AI Provider" description="Manage AI provider profiles and API keys." /><PageLoading /></div>
+
+  const selectedProfile = selectedSlug ? profiles[selectedSlug] : null
 
   return (
     <div className="flex flex-col flex-1 min-h-0">
-      <PageHeader title="AI Provider" description="Configure the AI backend, model, and API keys." />
-
-      {config ? (
+      <PageHeader title="AI Provider" description="Manage AI provider profiles and API keys." />
       <div className="flex-1 overflow-y-auto px-4 md:px-8 py-6">
-          <div className="max-w-[880px] mx-auto">
-            {/* Backend */}
-            <ConfigSection title="Backend" description="Changes take effect immediately.">
-              <div className="grid grid-cols-3 gap-3">
-                <BackendCard
-                  selected={uiBackend === 'agent-sdk'}
-                  onClick={() => handleBackendSwitch('agent-sdk')}
-                  icon={<svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><path d="M12 2a4 4 0 0 1 4 4v1a1 1 0 0 1-1 1H9a1 1 0 0 1-1-1V6a4 4 0 0 1 4-4z" /><path d="M8 8v2a4 4 0 0 0 8 0V8" /><path d="M12 14v4" /><path d="M8 22h8" /><circle cx="9" cy="5.5" r="0.5" fill="currentColor" stroke="none" /><circle cx="15" cy="5.5" r="0.5" fill="currentColor" stroke="none" /></svg>}
-                  title="Claude"
-                  description="Claude Code login or Anthropic API key"
-                />
-                <BackendCard
-                  selected={uiBackend === 'codex'}
-                  onClick={() => handleBackendSwitch('codex')}
-                  icon={<svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><polyline points="16 18 22 12 16 6" /><polyline points="8 6 2 12 8 18" /><line x1="14" y1="4" x2="10" y2="20" /></svg>}
-                  title="OpenAI / Codex"
-                  description="ChatGPT subscription or OpenAI API key"
-                />
-                <BackendCard
-                  selected={uiBackend === 'vercel-ai-sdk'}
-                  onClick={() => handleBackendSwitch('vercel-ai-sdk')}
-                  icon={<svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><path d="M13 10V3L4 14h7v7l9-11h-7z" /></svg>}
-                  title="Vercel AI SDK"
-                  description="Multi-provider, custom endpoints"
-                />
-              </div>
+        <div className="max-w-[880px] mx-auto">
+
+          {/* Profile List */}
+          <ConfigSection title="Profiles" description="Create multiple configurations and switch between them.">
+            <div className="space-y-2">
+              {Object.entries(profiles).map(([slug, profile]) => {
+                const info = BACKEND_INFO[profile.backend]
+                const isActive = slug === activeProfile
+                const isSelected = slug === selectedSlug
+                return (
+                  <button
+                    key={slug}
+                    onClick={() => { setSelectedSlug(slug); setCreating(null) }}
+                    className={`w-full flex items-center gap-3 p-3 rounded-lg border text-left transition-all ${
+                      isSelected
+                        ? 'border-accent bg-accent-dim/30'
+                        : 'border-border bg-bg hover:bg-bg-tertiary'
+                    }`}
+                  >
+                    <div className={`${isSelected ? 'text-accent' : 'text-text-muted'}`}>{info?.icon}</div>
+                    <div className="flex-1 min-w-0">
+                      <div className="flex items-center gap-2">
+                        <span className={`text-[13px] font-medium truncate ${isSelected ? 'text-accent' : 'text-text'}`}>{profile.label}</span>
+                        {isActive && <span className="text-[10px] px-1.5 py-0.5 rounded bg-accent/20 text-accent font-medium">Active</span>}
+                      </div>
+                      <p className="text-[11px] text-text-muted truncate">{info?.label} &middot; {profile.model}</p>
+                    </div>
+                  </button>
+                )
+              })}
+            </div>
+
+            {/* New Profile Button */}
+            <div className="mt-3 flex gap-2">
+              {(Object.keys(BACKEND_INFO) as AIBackend[]).map((backend) => (
+                <button
+                  key={backend}
+                  onClick={() => handleCreateStart(backend)}
+                  className={`flex items-center gap-1.5 px-3 py-1.5 rounded-lg text-[12px] font-medium border transition-all ${
+                    creating === backend
+                      ? 'border-accent text-accent bg-accent-dim/30'
+                      : 'border-border text-text-muted hover:text-text hover:bg-bg-tertiary'
+                  }`}
+                >
+                  <span>+</span> {BACKEND_INFO[backend].label}
+                </button>
+              ))}
+            </div>
+          </ConfigSection>
+
+          {/* Create Form */}
+          {creating && (
+            <ConfigSection title={`New ${BACKEND_INFO[creating].label} Profile`} description="Fill in the details and save.">
+              <ProfileForm
+                backend={creating}
+                onSave={handleCreateSave}
+                onCancel={() => setCreating(null)}
+              />
             </ConfigSection>
+          )}
 
-            {/* Auth mode (only for Agent SDK) */}
-            {uiBackend === 'agent-sdk' && (
-              <ConfigSection title="Authentication" description="Choose how Alice connects to Claude.">
-                <AgentSdkAuthForm aiProvider={config.aiProvider} onUpdate={(patch) => setConfig((c) => c ? { ...c, aiProvider: { ...c.aiProvider, ...patch } } : c)} />
-              </ConfigSection>
-            )}
+          {/* Edit Form */}
+          {selectedProfile && selectedSlug && !creating && (
+            <ConfigSection title={selectedProfile.label} description={`${BACKEND_INFO[selectedProfile.backend].label} profile — edit settings below.`}>
+              <ProfileEditor
+                slug={selectedSlug}
+                profile={selectedProfile}
+                isActive={selectedSlug === activeProfile}
+                onUpdate={(p) => handleProfileUpdate(selectedSlug, p)}
+                onSetActive={() => handleSetActive(selectedSlug)}
+                onDelete={() => handleDelete(selectedSlug)}
+              />
+            </ConfigSection>
+          )}
 
-            {/* Full model form (only for Vercel AI SDK) */}
-            {uiBackend === 'vercel-ai-sdk' && (
-              <ConfigSection title="Model" description="Provider, model, and API keys. Changes take effect on the next request.">
-                <ModelForm aiProvider={config.aiProvider} />
-              </ConfigSection>
-            )}
+          {/* Global API Keys */}
+          <ConfigSection title="Global API Keys" description="Shared across all profiles. Per-profile keys take priority.">
+            <ApiKeysForm currentStatus={apiKeys} onSaved={setApiKeys} />
+          </ConfigSection>
 
-            {/* Codex auth + model form */}
-            {uiBackend === 'codex' && (
-              <ConfigSection title="Authentication" description="Choose how Alice connects to OpenAI / Codex models.">
-                <CodexAuthForm aiProvider={config.aiProvider} onUpdate={(patch) => setConfig((c) => c ? { ...c, aiProvider: { ...c.aiProvider, ...patch } } : c)} />
-              </ConfigSection>
-            )}
-          </div>
+        </div>
       </div>
-      ) : (
-        <PageLoading />
-      )}
     </div>
   )
 }
 
-// ==================== OpenAI Form (simplified) ====================
+// ==================== Profile Form (Create) ====================
 
-function OpenAIForm({ aiProvider }: { aiProvider: AIProviderConfig }) {
-  const presets = PROVIDER_MODELS.openai
-  const initModel = aiProvider.provider === 'openai' && aiProvider.model ? aiProvider.model : presets[0].value
-  const isPreset = presets.some((p) => p.value === initModel)
-
-  const [model, setModel] = useState(isPreset ? initModel : '')
-  const [customModel, setCustomModel] = useState(isPreset ? '' : initModel)
-  const [baseUrl, setBaseUrl] = useState(aiProvider.baseUrl || '')
-  const [apiKey, setApiKey] = useState('')
-  const [keySaveStatus, setKeySaveStatus] = useState<SaveStatus>('idle')
-  const savedTimer = useRef<ReturnType<typeof setTimeout> | null>(null)
-
-  const effectiveModel = model || customModel
-
-  // Auto-save model + baseUrl
-  const modelData = useMemo(
-    () => ({
-      ...aiProvider,
-      provider: 'openai',
-      model: effectiveModel,
-      ...(baseUrl ? { baseUrl } : { baseUrl: undefined }),
-    }),
-    [aiProvider, effectiveModel, baseUrl],
-  )
-
-  const saveModel = useCallback(async (data: Record<string, unknown>) => {
-    await api.config.updateSection('aiProvider', data)
-  }, [])
-
-  const { status: modelStatus, retry: modelRetry } = useAutoSave({
-    data: modelData,
-    save: saveModel,
-  })
-
-  const hasKey = !!aiProvider.apiKeys?.openai
-
-  useEffect(() => () => { if (savedTimer.current) clearTimeout(savedTimer.current) }, [])
-
-  const handleSaveKey = async () => {
-    if (!apiKey) return
-    setKeySaveStatus('saving')
+function ProfileForm({ backend, onSave, onCancel }: {
+  backend: AIBackend
+  onSave: (slug: string, profile: Profile) => Promise<void>
+  onCancel: () => void
+}) {
+  const defaults = NEW_PROFILE_DEFAULTS[backend]
+  const [label, setLabel] = useState('')
+  const [model, setModel] = useState(defaults.model)
+  const [loginMethod, setLoginMethod] = useState(defaults.loginMethod ?? '')
+  const [provider, setProvider] = useState(defaults.provider ?? 'anthropic')
+  const [baseUrl, setBaseUrl] = useState('')
+  const [saving, setSaving] = useState(false)
+  const [error, setError] = useState('')
+
+  const handleSave = async () => {
+    if (!label.trim()) { setError('Label is required'); return }
+    setSaving(true)
+    setError('')
+    const slug = label.trim().toLowerCase().replace(/[^a-z0-9]+/g, '-').replace(/^-|-$/g, '')
+    if (!slug) { setError('Invalid label for slug generation'); setSaving(false); return }
+    const profile: Profile = {
+      backend,
+      label: label.trim(),
+      model,
+      ...(loginMethod ? { loginMethod } : {}),
+      ...(backend === 'vercel-ai-sdk' ? { provider } : {}),
+      ...(baseUrl ? { baseUrl } : {}),
+    }
     try {
-      const updatedKeys = { ...aiProvider.apiKeys, openai: apiKey }
-      await api.config.updateSection('aiProvider', { ...aiProvider, apiKeys: updatedKeys })
-      setApiKey('')
-      setKeySaveStatus('saved')
-      if (savedTimer.current) clearTimeout(savedTimer.current)
-      savedTimer.current = setTimeout(() => setKeySaveStatus('idle'), 2000)
-    } catch { setKeySaveStatus('error') }
-  }
-
-  const handleModelSelect = (value: string) => {
-    if (value === '__custom__') {
-      setModel('')
-      setCustomModel('')
-    } else {
-      setModel(value)
-      setCustomModel('')
+      await onSave(slug, profile)
+    } catch (err) {
+      setError(err instanceof Error ? err.message : 'Failed to save')
+    } finally {
+      setSaving(false)
     }
   }
 
   return (
-    <>
-      <Field label="Model">
-        <select
-          className={inputClass}
-          value={model || '__custom__'}
-          onChange={(e) => handleModelSelect(e.target.value)}
-        >
-          {presets.map((m) => (
-            <option key={m.value} value={m.value}>{m.label}</option>
-          ))}
-          <option value="__custom__">Custom...</option>
-        </select>
+    <div className="space-y-3">
+      <Field label="Label">
+        <input className={inputClass} value={label} onChange={(e) => setLabel(e.target.value)} placeholder="e.g. Claude Main, GPT Fast" />
       </Field>
-
-      {!model && (
-        <Field label="Custom Model ID">
-          <input
-            className={inputClass}
-            value={customModel}
-            onChange={(e) => setCustomModel(e.target.value)}
-            placeholder="e.g. gpt-4o, o3-pro"
-          />
-        </Field>
-      )}
-
-      <Field label="API Key">
-        <div className="relative">
-          <input
-            className={inputClass}
-            type="password"
-            value={apiKey}
-            onChange={(e) => setApiKey(e.target.value)}
-            placeholder={hasKey ? '(configured)' : 'sk-...'}
-          />
-          {hasKey && (
-            <span className="absolute right-2.5 top-1/2 -translate-y-1/2 text-[10px] text-green">active</span>
-          )}
-        </div>
-        <div className="flex items-center gap-3 mt-2">
-          <button
-            onClick={handleSaveKey}
-            disabled={!apiKey || keySaveStatus === 'saving'}
-            className="btn-primary"
-          >
-            Save Key
-          </button>
-          <SaveIndicator status={keySaveStatus} onRetry={handleSaveKey} />
-        </div>
-      </Field>
-
-      <Field label="Base URL" description="Leave empty for the official OpenAI API. Set for proxies or compatible endpoints (e.g. Azure OpenAI).">
-        <input
-          className={inputClass}
-          value={baseUrl}
-          onChange={(e) => setBaseUrl(e.target.value)}
-          placeholder="https://api.openai.com/v1"
-        />
-      </Field>
-
-      <SaveIndicator status={modelStatus} onRetry={modelRetry} />
-    </>
+      <ProfileFields backend={backend} model={model} setModel={setModel} loginMethod={loginMethod} setLoginMethod={setLoginMethod} provider={provider} setProvider={setProvider} baseUrl={baseUrl} setBaseUrl={setBaseUrl} />
+      {error && <p className="text-[12px] text-red">{error}</p>}
+      <div className="flex gap-2">
+        <button onClick={handleSave} disabled={saving} className="btn-primary">{saving ? 'Saving...' : 'Create Profile'}</button>
+        <button onClick={onCancel} className="btn-secondary">Cancel</button>
+      </div>
+    </div>
   )
 }
 
-// ==================== Model Form (full Vercel AI SDK) ====================
-
-function ModelForm({ aiProvider }: { aiProvider: AIProviderConfig }) {
-  // Detect whether saved config should render as "Custom" in the UI
-  const initCustom = detectCustomMode(aiProvider.provider || 'anthropic', aiProvider.model || '')
-  const [uiProvider, setUiProvider] = useState(initCustom ? 'custom' : (aiProvider.provider || 'anthropic'))
-  const [sdkProvider, setSdkProvider] = useState(aiProvider.provider || 'openai')
-  const [model, setModel] = useState(aiProvider.model || '')
-  const [customModel, setCustomModel] = useState(initCustom ? (aiProvider.model || '') : '')
-  const [baseUrl, setBaseUrl] = useState(aiProvider.baseUrl || '')
-  const [showKeys, setShowKeys] = useState(false)
-  const [keys, setKeys] = useState({ anthropic: '', openai: '', google: '' })
-  const [keySaveStatus, setKeySaveStatus] = useState<SaveStatus>('idle')
-  const keySavedTimer = useRef<ReturnType<typeof setTimeout> | null>(null)
-
-  const isCustomMode = uiProvider === 'custom'
-  const effectiveProvider = isCustomMode ? sdkProvider : uiProvider
-  const presets = PROVIDER_MODELS[uiProvider] || []
-  const isCustomModelInStandard = !isCustomMode && model !== '' && !presets.some((p) => p.value === model)
-  const effectiveModel = isCustomMode
-    ? customModel
-    : (isCustomModelInStandard ? customModel || model : model)
-
-  // Auto-save model/provider/baseUrl (but NOT apiKeys — those use manual save)
-  const modelData = useMemo(
-    () => ({
-      ...aiProvider,
-      provider: effectiveProvider,
-      model: effectiveModel,
-      ...(baseUrl ? { baseUrl } : { baseUrl: undefined }),
-    }),
-    [aiProvider, effectiveProvider, effectiveModel, baseUrl],
-  )
-
-  const saveModel = useCallback(async (data: Record<string, unknown>) => {
-    await api.config.updateSection('aiProvider', data)
-  }, [])
-
-  const { status: modelStatus, retry: modelRetry } = useAutoSave({
-    data: modelData,
-    save: saveModel,
-  })
-
-  // Derive key status from aiProvider config
-  const keyStatus = useMemo(() => ({
-    anthropic: !!aiProvider.apiKeys?.anthropic,
-    openai: !!aiProvider.apiKeys?.openai,
-    google: !!aiProvider.apiKeys?.google,
-  }), [aiProvider.apiKeys])
-
-  const [liveKeyStatus, setLiveKeyStatus] = useState(keyStatus)
+// ==================== Profile Editor (Edit existing) ====================
 
-  useEffect(() => setLiveKeyStatus(keyStatus), [keyStatus])
+function ProfileEditor({ slug, profile, isActive, onUpdate, onSetActive, onDelete }: {
+  slug: string
+  profile: Profile
+  isActive: boolean
+  onUpdate: (profile: Profile) => Promise<void>
+  onSetActive: () => void
+  onDelete: () => void
+}) {
+  const [label, setLabel] = useState(profile.label)
+  const [model, setModel] = useState(profile.model)
+  const [loginMethod, setLoginMethod] = useState(profile.loginMethod ?? '')
+  const [provider, setProvider] = useState(profile.provider ?? 'anthropic')
+  const [baseUrl, setBaseUrl] = useState(profile.baseUrl ?? '')
+  const [status, setStatus] = useState<SaveStatus>('idle')
+  const savedTimer = useRef<ReturnType<typeof setTimeout> | null>(null)
 
-  useEffect(() => () => {
-    if (keySavedTimer.current) clearTimeout(keySavedTimer.current)
-  }, [])
+  // Reset form when selected profile changes
+  useEffect(() => {
+    setLabel(profile.label)
+    setModel(profile.model)
+    setLoginMethod(profile.loginMethod ?? '')
+    setProvider(profile.provider ?? 'anthropic')
+    setBaseUrl(profile.baseUrl ?? '')
+    setStatus('idle')
+  }, [slug, profile])
 
-  const handleProviderChange = (newUiProvider: string) => {
-    setUiProvider(newUiProvider)
-    setBaseUrl('')
-    if (newUiProvider === 'custom') {
-      setSdkProvider('openai')
-      setModel('')
-      setCustomModel('')
-    } else {
-      setSdkProvider(newUiProvider)
-      const defaults = PROVIDER_MODELS[newUiProvider]
-      if (defaults?.length) {
-        setModel(defaults[0].value)
-        setCustomModel('')
-      } else {
-        setModel('')
-      }
-    }
-  }
+  useEffect(() => () => { if (savedTimer.current) clearTimeout(savedTimer.current) }, [])
 
-  const handleModelSelect = (value: string) => {
-    if (value === '__custom__') {
-      setModel('')
-      setCustomModel('')
-    } else {
-      setModel(value)
-      setCustomModel('')
+  const handleSave = async () => {
+    setStatus('saving')
+    const updated: Profile = {
+      backend: profile.backend,
+      label: label.trim() || profile.label,
+      model,
+      ...(loginMethod ? { loginMethod } : {}),
+      ...(profile.backend === 'vercel-ai-sdk' ? { provider } : {}),
+      ...(baseUrl ? { baseUrl } : {}),
+      ...(profile.apiKey ? { apiKey: profile.apiKey } : {}),
     }
-  }
-
-  const handleSaveKeys = async () => {
-    setKeySaveStatus('saving')
     try {
-      // Merge new keys into current aiProvider config
-      const updatedKeys = { ...aiProvider.apiKeys }
-      if (keys.anthropic) updatedKeys.anthropic = keys.anthropic
-      if (keys.openai) updatedKeys.openai = keys.openai
-      if (keys.google) updatedKeys.google = keys.google
-      await api.config.updateSection('aiProvider', { ...aiProvider, apiKeys: updatedKeys })
-      setLiveKeyStatus({
-        anthropic: !!updatedKeys.anthropic,
-        openai: !!updatedKeys.openai,
-        google: !!updatedKeys.google,
-      })
-      setKeys({ anthropic: '', openai: '', google: '' })
-      setKeySaveStatus('saved')
-      if (keySavedTimer.current) clearTimeout(keySavedTimer.current)
-      keySavedTimer.current = setTimeout(() => setKeySaveStatus('idle'), 2000)
+      await onUpdate(updated)
+      setStatus('saved')
+      if (savedTimer.current) clearTimeout(savedTimer.current)
+      savedTimer.current = setTimeout(() => setStatus('idle'), 2000)
     } catch {
-      setKeySaveStatus('error')
+      setStatus('error')
     }
   }
 
   return (
-    <>
-      <Field label="Provider">
-        <div className="flex border border-border rounded-lg overflow-hidden">
-          {PROVIDERS.map((p, i) => (
-            <button
-              key={p.value}
-              onClick={() => handleProviderChange(p.value)}
-              className={`flex-1 py-2 px-3 text-[13px] font-medium transition-colors ${
-                uiProvider === p.value
-                  ? 'bg-accent-dim text-accent'
-                  : 'bg-bg text-text-muted hover:bg-bg-tertiary hover:text-text'
-              } ${i > 0 ? 'border-l border-border' : ''}`}
-            >
-              {p.label}
-            </button>
-          ))}
-        </div>
+    <div className="space-y-3">
+      <Field label="Label">
+        <input className={inputClass} value={label} onChange={(e) => setLabel(e.target.value)} />
       </Field>
+      <ProfileFields backend={profile.backend} model={model} setModel={setModel} loginMethod={loginMethod} setLoginMethod={setLoginMethod} provider={provider} setProvider={setProvider} baseUrl={baseUrl} setBaseUrl={setBaseUrl} />
+      <div className="flex items-center gap-2 pt-1">
+        <button onClick={handleSave} className="btn-primary">Save Changes</button>
+        <SaveIndicator status={status} onRetry={handleSave} />
+        <div className="flex-1" />
+        {!isActive && (
+          <button onClick={onSetActive} className="text-[12px] text-accent hover:underline">Set as Default</button>
+        )}
+        {!isActive && (
+          <button onClick={onDelete} className="text-[12px] text-red hover:underline">Delete</button>
+        )}
+      </div>
+    </div>
+  )
+}
 
-      {/* Custom mode: API format selector */}
-      {isCustomMode && (
-        <Field label="API Format">
-          <select
-            className={inputClass}
-            value={sdkProvider}
-            onChange={(e) => setSdkProvider(e.target.value)}
-          >
-            {SDK_FORMATS.map((f) => (
-              <option key={f.value} value={f.value}>{f.label}</option>
-            ))}
-          </select>
-          <p className="text-[11px] text-text-muted mt-1">
-            Which API protocol does your endpoint speak?
-          </p>
-        </Field>
-      )}
+// ==================== Shared Profile Fields ====================
 
-      {/* Standard mode: preset model dropdown */}
-      {!isCustomMode && (
-        <Field label="Model">
-          <select
-            className={inputClass}
-            value={isCustomModelInStandard || model === '' ? '__custom__' : model}
-            onChange={(e) => handleModelSelect(e.target.value)}
-          >
-            {presets.map((m) => (
-              <option key={m.value} value={m.value}>{m.label}</option>
-            ))}
-            <option value="__custom__">Custom...</option>
+function ProfileFields({ backend, model, setModel, loginMethod, setLoginMethod, provider, setProvider, baseUrl, setBaseUrl }: {
+  backend: AIBackend
+  model: string; setModel: (v: string) => void
+  loginMethod: string; setLoginMethod: (v: string) => void
+  provider: string; setProvider: (v: string) => void
+  baseUrl: string; setBaseUrl: (v: string) => void
+}) {
+  return (
+    <>
+      {/* Login Method (agent-sdk and codex only) */}
+      {(backend === 'agent-sdk' || backend === 'codex') && (
+        <Field label="Authentication">
+          <select className={inputClass} value={loginMethod} onChange={(e) => setLoginMethod(e.target.value)}>
+            {backend === 'agent-sdk' ? (
+              <>
+                <option value="claudeai">Claude Pro/Max (subscription)</option>
+                <option value="api-key">API Key</option>
+              </>
+            ) : (
+              <>
+                <option value="codex-oauth">ChatGPT Subscription</option>
+                <option value="api-key">API Key</option>
+              </>
+            )}
           </select>
         </Field>
       )}
 
-      {/* Free-text model ID — always shown in custom mode, or when "Custom..." selected in standard mode */}
-      {(isCustomMode || isCustomModelInStandard || (!isCustomMode && model === '')) && (
-        <Field label={isCustomMode ? 'Model ID' : 'Custom Model ID'}>
-          <input
-            className={inputClass}
-            value={customModel || model}
-            onChange={(e) => { setCustomModel(e.target.value); setModel(e.target.value) }}
-            placeholder={isCustomMode ? 'e.g. gpt-4o, claude-3-opus' : 'e.g. claude-sonnet-4-5-20250929'}
-          />
+      {/* Provider (vercel-ai-sdk only) */}
+      {backend === 'vercel-ai-sdk' && (
+        <Field label="SDK Provider">
+          <select className={inputClass} value={provider} onChange={(e) => setProvider(e.target.value)}>
+            <option value="anthropic">Anthropic</option>
+            <option value="openai">OpenAI</option>
+            <option value="google">Google</option>
+          </select>
         </Field>
       )}
 
-      <Field label="Base URL">
-        <input
-          className={inputClass}
-          value={baseUrl}
-          onChange={(e) => setBaseUrl(e.target.value)}
-          placeholder={isCustomMode ? 'https://your-relay.example.com/v1' : 'Leave empty for official API'}
-        />
-        <p className="text-[11px] text-text-muted mt-1">
-          {isCustomMode ? 'Your relay or proxy endpoint.' : 'Custom endpoint for proxy or relay.'}
-        </p>
+      <Field label="Model">
+        <input className={inputClass} value={model} onChange={(e) => setModel(e.target.value)} placeholder="e.g. claude-sonnet-4-6, gpt-5.4" />
       </Field>
 
-      <SaveIndicator status={modelStatus} onRetry={modelRetry} />
-
-      {/* API Keys */}
-      <div className="mt-5 border-t border-border pt-4">
-        <button
-          onClick={() => setShowKeys(!showKeys)}
-          className="flex items-center gap-1.5 text-[13px] text-text-muted hover:text-text transition-colors"
-        >
-          <svg
-            width="10" height="10" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="2.5" strokeLinecap="round" strokeLinejoin="round"
-            className={`transition-transform ${showKeys ? 'rotate-90' : ''}`}
-          >
-            <polyline points="9 18 15 12 9 6" />
-          </svg>
-          API Keys
-          <span className="text-[11px] text-text-muted/60 ml-1">
-            ({Object.values(liveKeyStatus).filter(Boolean).length}/{Object.keys(liveKeyStatus).length} configured)
-          </span>
-        </button>
-
-        {showKeys && (
-          <div className="mt-3 space-y-3">
-            <p className="text-[11px] text-text-muted">
-              {isCustomMode
-                ? 'Enter the API key for your relay. It will be sent under the matching provider header.'
-                : 'Enter API keys below. Leave empty to keep existing value.'}
-            </p>
-            {(isCustomMode
-              ? SDK_FORMATS.filter((f) => f.value === sdkProvider)
-              : PROVIDERS.filter((p) => p.value !== 'custom')
-            ).map((p) => (
-              <Field key={p.value} label={isCustomMode ? `API Key (${p.label})` : `${p.label} API Key`}>
-                <div className="relative">
-                  <input
-                    className={inputClass}
-                    type="password"
-                    value={keys[p.value as keyof typeof keys] ?? ''}
-                    onChange={(e) => setKeys((k) => ({ ...k, [p.value]: e.target.value }))}
-                    placeholder={liveKeyStatus[p.value as keyof typeof liveKeyStatus] ? '(configured)' : 'Not configured'}
-                  />
-                  {liveKeyStatus[p.value as keyof typeof liveKeyStatus] && (
-                    <span className="absolute right-2.5 top-1/2 -translate-y-1/2 text-[10px] text-green">
-                      active
-                    </span>
-                  )}
-                </div>
-              </Field>
-            ))}
-            <div className="flex items-center gap-3">
-              <button
-                onClick={handleSaveKeys}
-                disabled={keySaveStatus === 'saving'}
-                className="btn-primary"
-              >
-                Save Keys
-              </button>
-              <SaveIndicator status={keySaveStatus} onRetry={handleSaveKeys} />
-            </div>
-          </div>
-        )}
-      </div>
+      <Field label="Base URL" description="Leave empty for official API. Set for proxies or compatible endpoints.">
+        <input className={inputClass} value={baseUrl} onChange={(e) => setBaseUrl(e.target.value)} placeholder="Leave empty for default" />
+      </Field>
     </>
   )
 }
 
-// ==================== Agent SDK Auth Form ====================
-
-const CODEX_LOGIN_METHODS: { value: LoginMethod; label: string; subtitle: string; hint: string }[] = [
-  { value: 'codex-oauth', label: 'ChatGPT Subscription', subtitle: 'Use your ChatGPT plan', hint: 'Requires local Codex CLI login (run codex login in terminal). Usage billed to your ChatGPT subscription.' },
-  { value: 'api-key', label: 'API Key', subtitle: 'Pay per token', hint: 'Enter your OpenAI API key below. Billed per token to your API account.' },
-]
+// ==================== Global API Keys ====================
 
-function CodexAuthForm({ aiProvider, onUpdate }: { aiProvider: AIProviderConfig; onUpdate: (patch: Partial<AIProviderConfig>) => void }) {
-  const [loginMethod, setLoginMethod] = useState<LoginMethod>(aiProvider.loginMethod ?? 'codex-oauth')
-  const [apiKey, setApiKey] = useState('')
-  const [keySaveStatus, setKeySaveStatus] = useState<SaveStatus>('idle')
+function ApiKeysForm({ currentStatus, onSaved }: {
+  currentStatus: Record<string, string | undefined>
+  onSaved: (status: Record<string, string | undefined>) => void
+}) {
+  const [keys, setKeys] = useState({ anthropic: '', openai: '', google: '' })
+  const [status, setStatus] = useState<SaveStatus>('idle')
   const savedTimer = useRef<ReturnType<typeof setTimeout> | null>(null)
 
   useEffect(() => () => { if (savedTimer.current) clearTimeout(savedTimer.current) }, [])
 
-  const handleLoginMethodChange = async (method: LoginMethod) => {
-    setLoginMethod(method)
-    try {
-      await api.config.updateSection('aiProvider', { ...aiProvider, loginMethod: method })
-      onUpdate({ loginMethod: method })
-    } catch { setLoginMethod(loginMethod) }
-  }
-
-  const handleSaveKey = async () => {
-    if (!apiKey) return
-    setKeySaveStatus('saving')
+  const handleSave = async () => {
+    setStatus('saving')
     try {
-      const updatedKeys = { ...aiProvider.apiKeys, openai: apiKey }
-      await api.config.updateSection('aiProvider', { ...aiProvider, apiKeys: updatedKeys })
-      onUpdate({ apiKeys: updatedKeys })
-      setApiKey('')
-      setKeySaveStatus('saved')
+      const toSave: Record<string, string> = {}
+      if (keys.anthropic) toSave.anthropic = keys.anthropic
+      if (keys.openai) toSave.openai = keys.openai
+      if (keys.google) toSave.google = keys.google
+      await api.config.updateApiKeys(toSave)
+      onSaved({
+        ...currentStatus,
+        ...(keys.anthropic ? { anthropic: '(set)' } : {}),
+        ...(keys.openai ? { openai: '(set)' } : {}),
+        ...(keys.google ? { google: '(set)' } : {}),
+      })
+      setKeys({ anthropic: '', openai: '', google: '' })
+      setStatus('saved')
       if (savedTimer.current) clearTimeout(savedTimer.current)
-      savedTimer.current = setTimeout(() => setKeySaveStatus('idle'), 2000)
-    } catch { setKeySaveStatus('error') }
-  }
-
-  return (
-    <>
-      <div className="grid grid-cols-2 gap-3">
-        {CODEX_LOGIN_METHODS.map((m) => (
-          <BackendCard
-            key={m.value}
-            selected={loginMethod === m.value}
-            onClick={() => handleLoginMethodChange(m.value)}
-            icon={m.value === 'codex-oauth'
-              ? <svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><path d="M15 3h4a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2h-4" /><polyline points="10 17 15 12 10 7" /><line x1="15" y1="12" x2="3" y2="12" /></svg>
-              : <svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><rect x="3" y="11" width="18" height="11" rx="2" ry="2" /><path d="M7 11V7a5 5 0 0 1 10 0v4" /></svg>}
-            title={m.label}
-            description={m.subtitle}
-          />
-        ))}
-      </div>
-      <p className="text-[11px] text-text-muted mt-2">
-        {CODEX_LOGIN_METHODS.find((m) => m.value === loginMethod)?.hint}
-      </p>
-
-      {loginMethod === 'api-key' && (
-        <Field label="OpenAI API Key">
-          <div className="relative">
-            <input
-              className={inputClass}
-              type="password"
-              value={apiKey}
-              onChange={(e) => setApiKey(e.target.value)}
-              placeholder={aiProvider.apiKeys?.openai ? '(configured)' : 'sk-...'}
-            />
-            {aiProvider.apiKeys?.openai && (
-              <span className="absolute right-2.5 top-1/2 -translate-y-1/2 text-[10px] text-green">active</span>
-            )}
-          </div>
-          <div className="flex items-center gap-3 mt-2">
-            <button
-              onClick={handleSaveKey}
-              disabled={!apiKey || keySaveStatus === 'saving'}
-              className="btn-primary"
-            >
-              Save Key
-            </button>
-            <SaveIndicator status={keySaveStatus} onRetry={handleSaveKey} />
-          </div>
-        </Field>
-      )}
-    </>
-  )
-}
-
-function AgentSdkAuthForm({ aiProvider, onUpdate }: { aiProvider: AIProviderConfig; onUpdate: (patch: Partial<AIProviderConfig>) => void }) {
-  const [loginMethod, setLoginMethod] = useState<LoginMethod>(aiProvider.loginMethod ?? 'api-key')
-  const [apiKey, setApiKey] = useState('')
-  const [keySaveStatus, setKeySaveStatus] = useState<SaveStatus>('idle')
-  const savedTimer = useRef<ReturnType<typeof setTimeout> | null>(null)
-
-  useEffect(() => () => { if (savedTimer.current) clearTimeout(savedTimer.current) }, [])
-
-  const handleLoginMethodChange = async (method: LoginMethod) => {
-    setLoginMethod(method)
-    try {
-      await api.config.updateSection('aiProvider', { ...aiProvider, loginMethod: method })
-      onUpdate({ loginMethod: method })
-    } catch { /* revert on failure */ setLoginMethod(loginMethod) }
+      savedTimer.current = setTimeout(() => setStatus('idle'), 2000)
+    } catch {
+      setStatus('error')
+    }
   }
 
-  const handleSaveKey = async () => {
-    if (!apiKey) return
-    setKeySaveStatus('saving')
-    try {
-      const updatedKeys = { ...aiProvider.apiKeys, anthropic: apiKey }
-      await api.config.updateSection('aiProvider', { ...aiProvider, apiKeys: updatedKeys })
-      onUpdate({ apiKeys: updatedKeys })
-      setApiKey('')
-      setKeySaveStatus('saved')
-      if (savedTimer.current) clearTimeout(savedTimer.current)
-      savedTimer.current = setTimeout(() => setKeySaveStatus('idle'), 2000)
-    } catch { setKeySaveStatus('error') }
-  }
+  const fields = [
+    { key: 'anthropic', label: 'Anthropic', placeholder: 'sk-ant-...' },
+    { key: 'openai', label: 'OpenAI', placeholder: 'sk-...' },
+    { key: 'google', label: 'Google', placeholder: 'AIza...' },
+  ] as const
 
   return (
     <>
-      <div className="grid grid-cols-2 gap-3">
-        {LOGIN_METHODS.map((m) => (
-          <BackendCard
-            key={m.value}
-            selected={loginMethod === m.value}
-            onClick={() => handleLoginMethodChange(m.value)}
-            icon={m.value === 'claudeai'
-              ? <svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><path d="M15 3h4a2 2 0 0 1 2 2v14a2 2 0 0 1-2 2h-4" /><polyline points="10 17 15 12 10 7" /><line x1="15" y1="12" x2="3" y2="12" /></svg>
-              : <svg width="18" height="18" viewBox="0 0 24 24" fill="none" stroke="currentColor" strokeWidth="1.5" strokeLinecap="round" strokeLinejoin="round"><rect x="3" y="11" width="18" height="11" rx="2" ry="2" /><path d="M7 11V7a5 5 0 0 1 10 0v4" /></svg>}
-            title={m.label}
-            description={m.subtitle}
-          />
-        ))}
-      </div>
-      <p className="text-[11px] text-text-muted mt-2">
-        {LOGIN_METHODS.find((m) => m.value === loginMethod)?.hint}
-      </p>
-
-      {loginMethod === 'api-key' && (
-        <Field label="Anthropic API Key">
+      {fields.map((f) => (
+        <Field key={f.key} label={`${f.label} API Key`}>
           <div className="relative">
             <input
               className={inputClass}
               type="password"
-              value={apiKey}
-              onChange={(e) => setApiKey(e.target.value)}
-              placeholder={aiProvider.apiKeys?.anthropic ? '(configured)' : 'sk-ant-...'}
+              value={keys[f.key]}
+              onChange={(e) => setKeys((k) => ({ ...k, [f.key]: e.target.value }))}
+              placeholder={currentStatus[f.key] ? '(configured)' : f.placeholder}
             />
-            {aiProvider.apiKeys?.anthropic && (
+            {currentStatus[f.key] && (
               <span className="absolute right-2.5 top-1/2 -translate-y-1/2 text-[10px] text-green">active</span>
             )}
           </div>
-          <div className="flex items-center gap-3 mt-2">
-            <button
-              onClick={handleSaveKey}
-              disabled={!apiKey || keySaveStatus === 'saving'}
-              className="btn-primary"
-            >
-              Save Key
-            </button>
-            <SaveIndicator status={keySaveStatus} onRetry={handleSaveKey} />
-          </div>
         </Field>
-      )}
+      ))}
+      <div className="flex items-center gap-3">
+        <button onClick={handleSave} disabled={status === 'saving'} className="btn-primary">Save Keys</button>
+        <SaveIndicator status={status} onRetry={handleSave} />
+      </div>
     </>
   )
 }

From a9d2dba4e14340d446de050bd5ddd63d5cb333d0 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 21:04:52 +0800
Subject: [PATCH 13/15] fix: add store:false and array input for Codex
 subscription endpoint

ChatGPT's /backend-api/codex/responses requires:
- store: false (server-side storage not supported on subscription endpoint)
- input as array (not string)

Without these, requests get 400 with "Store must be set to false" or
"Input must be a list".

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 src/ai-providers/codex/codex-provider.ts | 31 +++++++++++++++++-------
 1 file changed, 22 insertions(+), 9 deletions(-)

diff --git a/src/ai-providers/codex/codex-provider.ts b/src/ai-providers/codex/codex-provider.ts
index e13af0da..9cd4aef9 100644
--- a/src/ai-providers/codex/codex-provider.ts
+++ b/src/ai-providers/codex/codex-provider.ts
@@ -73,7 +73,8 @@ export class CodexProvider implements AIProvider {
       const response = await client.responses.create({
         model,
         instructions,
-        input: prompt,
+        input: [{ role: 'user' as const, content: prompt }],
+        store: false,
       })
 
       const text = response.output
@@ -145,6 +146,7 @@ export class CodexProvider implements AIProvider {
           instructions,
           input,
           tools: tools.length > 0 ? tools : undefined,
+          store: false, // Required by ChatGPT subscription endpoint
         })
 
         for await (const event of stream) {
@@ -168,15 +170,26 @@ export class CodexProvider implements AIProvider {
           yield { type: 'done', result: { text: errorText, media: [] } }
           return
         }
-        logger.error({
-          err: err?.message,
+        // Extract all available error detail from OpenAI SDK error
+        const errDetail: Record<string, unknown> = {
+          message: err?.message,
           status: err?.status,
-          headers: err?.headers,
-          body: err?.error ?? err?.body,
-          model,
-          inputItems: input.length,
-          toolCount: tools.length,
-        }, 'responses_api_error')
+          type: err?.type,
+          code: err?.code,
+          param: err?.param,
+        }
+        // The SDK stores the parsed error body in err.error
+        if (err?.error) errDetail.errorBody = err.error
+        // Raw response headers can help debug
+        if (err?.headers) {
+          const h: Record<string, string> = {}
+          try { for (const [k, v] of Object.entries(err.headers)) h[k] = String(v) } catch {}
+          if (Object.keys(h).length > 0) errDetail.headers = h
+        }
+        errDetail.model = model
+        errDetail.inputItems = input.length
+        errDetail.toolCount = tools.length
+        logger.error(errDetail, 'responses_api_error')
         const errorText = accumulatedText + stepText +
           `\n\n[Codex API error: ${err?.message ?? 'unknown error'}]`
         yield { type: 'done', result: { text: errorText, media: [] } }

From 49fb3363a57f99a3610c41593624993ac4610ef0 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 21:11:03 +0800
Subject: [PATCH 14/15] refactor: Codex provider uses structured input instead
 of text history

Replace toTextHistory() + buildChatHistoryPrompt() with new
toResponsesInput() that converts SessionEntry[] to Responses API
input items (messages, function_call, function_call_output).

The model now sees proper multi-turn conversation structure instead of
a single user message with XML-wrapped history text. Handles orphaned
tool calls from compaction truncation.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 src/ai-providers/codex/codex-provider.ts | 13 ++---
 src/core/session.ts                      | 70 ++++++++++++++++++++++++
 2 files changed, 75 insertions(+), 8 deletions(-)

diff --git a/src/ai-providers/codex/codex-provider.ts b/src/ai-providers/codex/codex-provider.ts
index 9cd4aef9..96e82d1a 100644
--- a/src/ai-providers/codex/codex-provider.ts
+++ b/src/ai-providers/codex/codex-provider.ts
@@ -16,8 +16,7 @@ import { pino } from 'pino'
 import type { ProviderResult, ProviderEvent, AIProvider, GenerateOpts } from '../types.js'
 import type { SessionEntry } from '../../core/session.js'
 import type { ResolvedProfile } from '../../core/config.js'
-import { toTextHistory } from '../../core/session.js'
-import { buildChatHistoryPrompt, DEFAULT_MAX_HISTORY } from '../utils.js'
+import { toResponsesInput } from '../../core/session.js'
 import { readAgentConfig } from '../../core/config.js'
 import { getAccessToken, clearTokenCache } from './auth.js'
 import { convertTools } from './tool-bridge.js'
@@ -96,10 +95,6 @@ export class CodexProvider implements AIProvider {
     prompt: string,
     opts?: GenerateOpts,
   ): AsyncGenerator<ProviderEvent> {
-    const maxHistory = opts?.maxHistoryEntries ?? DEFAULT_MAX_HISTORY
-    const textHistory = toTextHistory(entries).slice(-maxHistory)
-    const fullPrompt = buildChatHistoryPrompt(prompt, textHistory, opts?.historyPreamble)
-
     const { client, model } = await this.createClient(opts?.profile)
     const instructions = opts?.systemPrompt ?? await this.getSystemPrompt()
     const agentConfig = await readAgentConfig()
@@ -109,9 +104,11 @@ export class CodexProvider implements AIProvider {
     const allTools = await this.getTools()
     const tools = convertTools(allTools, opts?.disabledTools)
 
-    // Build initial input
+    // Build structured input from session history + current prompt
+    const history = toResponsesInput(entries)
     const input: OpenAI.Responses.ResponseInputItem[] = [
-      { role: 'user', content: fullPrompt, type: 'message' },
+      ...history as OpenAI.Responses.ResponseInputItem[],
+      { role: 'user', content: prompt },
     ]
 
     yield* this.toolLoop(client, model, instructions, input, tools, allTools, maxSteps)
diff --git a/src/core/session.ts b/src/core/session.ts
index b0a94444..2b0484c2 100644
--- a/src/core/session.ts
+++ b/src/core/session.ts
@@ -466,6 +466,76 @@ export function toTextHistory(entries: SessionEntry[]): Array<{ role: 'user' | '
   return history
 }
 
+// ==================== Responses API Input (for Codex provider) ====================
+
+/**
+ * Input item types for OpenAI's Responses API.
+ * Mirrors the subset of ResponseInputItem that we actually use.
+ */
+export type ResponsesInputItem =
+  | { role: 'user' | 'assistant'; content: string }
+  | { type: 'function_call'; call_id: string; name: string; arguments: string }
+  | { type: 'function_call_output'; call_id: string; output: string }
+
+/**
+ * Convert session entries → OpenAI Responses API input items.
+ *
+ * Similar to toModelMessages() but targets the Responses API format.
+ * Handles orphaned tool calls (compaction truncation) by stripping
+ * function_call items that have no matching function_call_output.
+ */
+export function toResponsesInput(entries: SessionEntry[]): ResponsesInputItem[] {
+  const items: ResponsesInputItem[] = []
+
+  for (const entry of entries) {
+    if (entry.type === 'system' && entry.subtype === 'compact_boundary') continue
+
+    const { message } = entry
+
+    if (message.role === 'user') {
+      if (typeof message.content === 'string') {
+        items.push({ role: 'user', content: message.content })
+      } else {
+        // tool_result blocks → function_call_output items
+        for (const block of message.content) {
+          if (block.type === 'tool_result') {
+            items.push({ type: 'function_call_output', call_id: block.tool_use_id, output: block.content })
+          } else if (block.type === 'text') {
+            items.push({ role: 'user', content: block.text })
+          }
+        }
+      }
+    } else if (message.role === 'assistant') {
+      if (typeof message.content === 'string') {
+        items.push({ role: 'assistant', content: message.content })
+      } else {
+        for (const block of message.content) {
+          if (block.type === 'text') {
+            items.push({ role: 'assistant', content: block.text })
+          } else if (block.type === 'tool_use') {
+            items.push({
+              type: 'function_call',
+              call_id: block.id,
+              name: block.name,
+              arguments: typeof block.input === 'string' ? block.input : JSON.stringify(block.input),
+            })
+          }
+        }
+      }
+    }
+  }
+
+  // Sanitize: strip function_call items without a matching function_call_output
+  const outputIds = new Set<string>()
+  for (const item of items) {
+    if ('type' in item && item.type === 'function_call_output') outputIds.add(item.call_id)
+  }
+  return items.filter((item) => {
+    if ('type' in item && item.type === 'function_call') return outputIds.has(item.call_id)
+    return true
+  })
+}
+
 // ==================== Chat History (for Web UI) ====================
 
 /** A display-ready chat history item — either plain text or a group of paired tool calls. */

From d3ce583f510e19f12cbb6614e20b3a89988e8419 Mon Sep 17 00:00:00 2001
From: Ame <ame@AmedeMacBook-Pro.local>
Date: Tue, 7 Apr 2026 21:21:12 +0800
Subject: [PATCH 15/15] fix: use output_item.done for function call_id instead
 of arguments.done

The function_call_arguments.done event lacks call_id and name fields.
The complete function call object (with call_id, name, arguments) is
only available on the response.output_item.done event.

Verified with streaming round-trip test against live Codex endpoint.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---
 src/ai-providers/codex/codex-provider.ts | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)

diff --git a/src/ai-providers/codex/codex-provider.ts b/src/ai-providers/codex/codex-provider.ts
index 96e82d1a..3cacde5f 100644
--- a/src/ai-providers/codex/codex-provider.ts
+++ b/src/ai-providers/codex/codex-provider.ts
@@ -150,12 +150,17 @@ export class CodexProvider implements AIProvider {
           if (event.type === 'response.output_text.delta') {
             yield { type: 'text', text: event.delta }
             stepText += event.delta
-          } else if (event.type === 'response.function_call_arguments.done') {
-            functionCalls.push({
-              call_id: event.item_id,
-              name: event.name,
-              arguments: event.arguments,
-            })
+          } else if (event.type === 'response.output_item.done') {
+            // function_call_arguments.done lacks call_id and name;
+            // output_item.done carries the complete function call object.
+            const item = (event as any).item
+            if (item?.type === 'function_call') {
+              functionCalls.push({
+                call_id: item.call_id,
+                name: item.name,
+                arguments: item.arguments,
+              })
+            }
           }
         }
       } catch (err: any) {