System prompt now explicitly:
- Lists which assets have real-time data (crypto) and which don't (stocks/forex)
- FORBIDS fabricating specific prices
- Requires disclosure: 'based on historical knowledge, not live data'
- Allows strategy frameworks but not fake price levels
Tested: 拓维信息 → now says '我没有A股实时数据' upfront, gives
strategy framework without fake prices. Honesty > confidence.
SSRF protection was blocking AI API calls because Go's custom resolver
couldn't resolve external domains (DNS timeout). curl worked because it
uses the system resolver.
Fix: Agent creates MCP client with standard http.Client (no SSRF filter)
since we control the API URLs.
Tested: '帮我看看拓维信息这只股，我有10万元，帮我制定1个月交易计划'
→ Returns professional A-share analysis with entry/exit strategy in 5.8s
DELETED: agent/router.go (regex-based intent routing)
The fundamental architecture was wrong. A real AI Agent doesn't use
pattern matching to understand users. It uses AI.
New flow:
1. User says anything
2. /help and /status → instant response (no AI needed)
3. Setup keywords → onboard flow
4. EVERYTHING else → LLM with full context
The LLM receives:
- System prompt with NOFXi's capabilities and behavior rules
- Real-time market data (auto-detected from mentioned symbols)
- Live trader positions and status
- User's message
The LLM decides what to do. Not regex. Not if-else. The AI.
This is what makes it an Agent, not a chatbot.
1. Complex natural language → AI chat (not regex analyze)
'拓维信息这个股怎么样帮我写策略' now goes to AI, not broken regex
2. Agent auto-loads AI client from store on startup
- Scans AIModel configs, creates MCP client from first available
- No manual injection needed
3. Added start.sh for one-command startup (backend + frontend)
Routing logic:
- /analyze BTC (simple, ≤3 words) → structured analyze
- Complex sentences → AI chat with full context
- This is the key insight: regex can't parse natural language, AI can
- Crypto: real-time data (market.Get) + AI analysis
- Stocks/Forex: AI answers with its own knowledge (A股/港股/美股/外汇)
- No more 'coming soon' — if AI is configured, it can analyze anything
- Graceful fallback: no AI → guide user to configure
- AI prompt asks for: asset type, fundamentals, trend, key levels, risks
Core behavior change:
- Agent NO LONGER auto-prompts setup on first message
- Setup only triggers on explicit: '开始配置' / 'setup' / '绑定交易所'
- Normal questions answered normally, even without trader configured
- AI analysis works without AI model (shows raw data + guidance)
- Stock analysis shows 'coming soon' instead of error
Intent routing improved:
- '你能分析拓维信息吗' → correctly routes to analyze
- Chinese suffixes cleaned (吗/呢/这只股票)
- More natural language patterns matched
Philosophy: Agent should be helpful from minute one.
Configuration is optional, not a gate.
- Detects when user sends Chinese text, short text, or text with spaces
during key-input steps → auto-pauses setup and answers the question
- API key masking fixed (no more garbled display)
- User can resume setup with '开始配置' anytime
Key changes:
- Agent API routes (/api/agent/*) registered directly on NOFX's :8080 server
- No more separate :8900 port — everything on :8080
- AgentChatPage.tsx added to React frontend
- Agent tab (🤖 Agent) added to navigation bar, set as DEFAULT page
- Branding: NOFX → NOFXi everywhere (title, header, branding.ts, i18n)
- Default landing page is now Agent chat, not competition
This is THE change: there is no more 'NOFX' vs 'NOFXi'.
Users open localhost:8080, they see NOFXi.
Agent chat is the first thing they see.
All other pages (Traders, Dashboard, Strategy, Data) are still there.
One binary. One port. One product. NOFXi.
Users no longer need Web UI to configure NOFXi. The agent guides them
through setup via natural language:
1. First message → agent detects no traders → shows welcome
2. User says '开始配置' or 'setup'
3. Agent asks: which exchange? (1-7, name, or Chinese name)
4. Agent asks: API Key → API Secret → Passphrase (if needed)
5. Agent asks: which AI model? (DeepSeek/Qwen/GPT/Claude/Skip)
6. Agent asks: AI API Key
7. Agent creates Exchange + AIModel + Trader in NOFX store
8. Done. User can immediately start using /analyze, /positions, etc.
Features:
- Full Chinese/English bilingual flow
- Cancel anytime with '取消' or 'cancel'
- State persisted in SystemConfig (survives restarts)
- Supports all 7 exchanges + 4 AI providers
- OKX/Bitget/KuCoin auto-asks for Passphrase
- API keys stored encrypted via NOFX crypto service
Critical fixes:
1. Agent now uses NOFX's MCP client for AI calls (real models, not placeholder)
- Gets AI client from first configured trader's MCP config
- Added AutoTrader.GetMCPClient() export
2. /analyze uses REAL market data (market.Get → EMA/MACD/RSI/BB/OI/funding)
then sends to AI — no more hallucinated prices
3. AI chat enriches prompts with live market data when user mentions a coin
4. Fallback: if AI unavailable, shows raw formatted market data (still useful)
Router improvements:
- Better natural language detection for Chinese
- '该不该买' '走势' '趋势' '行情' → analyze intent
- '多少钱' '赚了' '亏了' → query intent
- Fewer messages falling through to generic chat
Web UI:
- Endpoints updated to /api/agent/* namespace
- nofxi.html served at agent root (localhost:8900)
Result: AI analysis now shows real BTC price, real RSI, real EMA — not fiction.
This is the key architectural change: NOFXi is no longer a separate
sub-project in nofxi/, but a first-class module in agent/ that sits
on top of NOFX's existing engine.
agent/ package:
- agent.go: Core orchestrator, directly uses NOFX's TraderManager,
Store, and Market packages. No wrapper/bridge needed.
- brain.go: Proactive intelligence (news scan, market briefs, signal handling)
- sentinel.go: Market anomaly detection (price breakout, volume spike, funding rate)
- scheduler.go: Daily reports, position risk checks
- router.go: Intent routing for natural language (中/英 bilingual)
- i18n.go: Full Chinese/English message catalog
- web.go: Agent REST API on :8900 (chat, klines, ticker)
main.go changes:
- Banner: NOFX → NOFXi
- Agent auto-starts with NOFX (sentinel + brain + scheduler + web)
- Zero breaking changes to existing NOFX functionality
Key difference from old nofxi/:
- Agent reads LIVE data from TraderManager (positions, balances, status)
- Uses market.Get() for full technical indicators (EMA/MACD/RSI/BB)
- No duplicate code — kernel, trader, store all used directly
- Old nofxi/ sub-project will be removed once migration is verified
Architecture:
NOFX (engine) → kernel + trader + market + store + mcp + telegram
NOFXi (brain) → agent/ package → uses all of the above
Full audit of NOFX's capabilities vs NOFXi's current integration status.
Priority roadmap: P0 (kernel+trader+market+mcp) → P1 (store+risk) → P2 (grid+web)
Agent i18n:
- All command responses bilingual (help, status, positions, balance, trades, analysis, settings, errors)
- Language auto-detected from web UI lang toggle
- Stored per-user in SQLite preferences
- AI chat system prompt switches to Chinese/English based on user lang
- Trade confirmation prompts bilingual
Web UI:
- Sends lang parameter with every API request
- Agent extracts [lang:xx] prefix and persists preference
i18n message catalog: nofxi/internal/agent/i18n.go
- Track consecutive AI failures
- After 3 failures: activate safe mode (no new positions, close/hold only)
- Auto-deactivate when AI recovers
- Keep loop running in safe mode (retry each cycle)
- Log clearly: 🛡️ SAFE MODE ACTIVATED/DEACTIVATED
sonner's toast.promise() returns a toast ID (not a Promise), so
await resolves immediately and subsequent data refresh fetches stale
data. Users had to manually reload to see changes.
Fixed across AITradersPage, SettingsPage, and TraderConfigModal.
* feat: add X-Client-ID header for claw402 monitoring
* feat(mcp): add context length guard to prevent oversized requests
- Add MaxContext field to Config (default 0 = no limit)
- Add WithMaxContext() option for setting model context limits
- Add context_guard.go: token estimation + message truncation
- Integrate guard into both BuildMCPRequestBody and BuildRequestBodyFromRequest
- Support both map[string]string and map[string]any message formats
- Truncates oldest non-system messages when estimated tokens exceed limit
- Always preserves system messages and keeps at least 1 non-system message
- Logs warning when truncation occurs for debugging
Usage: mcp.NewDeepSeekClient(mcp.WithMaxContext(131072))
- Extract ParseSSEStream as shared function from CallWithRequestStream
- Add DoX402RequestStream and X402CallStream for streaming x402 payments
- Switch Claw402Client.Call to use streaming (X402CallStream)
- TeeReader fallback: SSE parsing with JSON fallback for non-SSE responses
- Idle timeout watchdog (90s) protects against stalled streams
CoinSourceEditor constructs keys as `${value}Desc` where value='oi_top',
expecting 'oi_topDesc' but the translation key was 'oiTopDesc' (camelCase).
This caused ts(undefined, lang) → "Cannot read properties of undefined".
- Add missing configResponse.ok check in handleCreateStrategy to prevent
rendering with invalid config data when API fails
- Remove deleted benefit3 translation key reference from LoginRequiredOverlay
The outer retry loop in client.go re-initiates the entire x402 payment
flow on each attempt, causing duplicate USDC charges. The inner x402
retry loop (5 attempts with re-signing) already handles all retryable
scenarios. Set MaxRetries=1 for Claw402, BlockRunBase, and BlockRunSol
to ensure only one payment per AI decision.
AI inference (especially DeepSeek) often exceeds the default 120s HTTP
timeout, causing the client to disconnect while the server completes
successfully — resulting in repeated payments on each retry.
Changes:
- Set X402Timeout = 5min for all x402 providers (Claw402, BlockRunBase, BlockRunSol)
- Handle 402 during payment retry by re-extracting Payment-Required
header and re-signing instead of failing immediately
- Increase payment retry attempts from 3 to 5 for unstable gateways
The .gitignore had a Python-inherited `lib/` rule that blocked the
entire web/src/lib/api/ directory from being tracked. The original
api.ts was a file (not matched), but the split created a directory
inside lib/ which was ignored. Remove the `lib/` gitignore rule and
add the api module split files that were missing from the previous commit.
Add retry loop (up to 3 attempts with exponential backoff) for 5xx
server errors on payment-signed x402 requests, reusing the same
payment signature to avoid double-charges. Also add 502/503/520/524
to the retryable error patterns in the MCP client.
Remove the entire AI Debate Arena module (~5,300 lines) to simplify
the codebase. This removes the multi-AI debate trading decision system
including backend engine, API handlers, database store, frontend page,
navigation, translations, and documentation references.
* feat(telegram): add AI agent bot with streaming and account context
- Add Telegram bot with long-polling and AI agent loop (api_call tool)
- SSE streaming with real-time message editing and ⏳ placeholder
- Account state injection at conversation start (models, exchanges,
strategies, traders, per-trader PnL and statistics)
- Lane semaphore per chat serializes concurrent messages (60s timeout)
- Idle timeout watchdog (60s) prevents hung streaming connections
- Look-ahead buffer prevents partial <api_call> tag leaking to user
- Fix PUT /strategies/:id to merge config (read-then-merge pattern)
- Add route registry with full API schema for LLM documentation
- Add TelegramConfig store and Web UI config modal
- Add GetAnyEnabled to AIModel store for bot LLM client selection
* fix(telegram): eliminate narration, add full-setup workflow and tests
- Rewrite NO NARRATION rule: response is EITHER api_call tag alone OR
final text reply — no text before api_call under any circumstances
- Ban all narration patterns: 现在我将/好的/正在/I will/Let me etc.
- Add 'create strategy + create trader + start' full setup workflow
- Add 12 automated tests covering:
- No narration leaking to user (5 narration variants tested)
- api_call tag never leaks to user
- Full setup workflow: POST strategy → verify → POST trader → start
- Start existing trader workflow
- Max iterations safety, tag stripping, parser edge cases
* refactor(agent): replace XML api_call with native function calling
Migrate the Telegram bot agent from an XML tag hack (<api_call>) to
OpenAI-native function calling via CallWithRequestFull.
Key changes:
- mcp/interface.go: add parseMCPResponseFull to clientHooks interface
- mcp/client.go: route callWithRequestFull through hooks for overridability
- mcp/claude_client.go: override parseMCPResponseFull for Claude response
format (tool_use blocks instead of choices[].message.tool_calls)
- telegram/agent/agent.go: rewrite Run() to use CallWithRequestFull;
define api_request tool with JSON Schema; implement tool-call loop
with role="tool" result messages; remove XML parsing entirely
- telegram/agent/apicall.go: remove parseAPICall (dead code)
- telegram/agent/prompt.go: simplify — remove XML format instructions,
replace with concise api_request tool usage instructions
- telegram/agent/agent_test.go: rebuild all tests using LLMResponse
objects; add TestNarrationStructurallyImpossible, TestOnChunkCalledWithFinalReply,
TestToolCallIDPropagated; remove XML-specific tests
Architecture advantage: with native function calling, the LLM returns
EITHER ToolCalls OR Content — never both. Narration is now structurally
impossible at the protocol level, not just enforced by prompt rules.
All 11 agent tests pass. mcp package tests pass.
* refactor(mcp): route buildRequestBodyFromRequest through hooks + full Anthropic format
Problem: callWithRequest/Full/Stream all called client.buildRequestBodyFromRequest
directly (not via hooks), so ClaudeClient could never override it. This meant
tool calling sent OpenAI format to Anthropic (wrong field names, wrong roles).
Changes:
mcp/interface.go
- Add buildRequestBodyFromRequest(*Request) map[string]any to clientHooks
- Improve comments: document what each hook group does and why
mcp/client.go
- All three paths (callWithRequest, callWithRequestFull, CallWithRequestStream)
now call client.hooks.buildRequestBodyFromRequest — ClaudeClient picks up
mcp/claude_client.go
- Full rewrite with format comparison table in package doc
- buildRequestBodyFromRequest: produces correct Anthropic wire format
* system prompt → top-level "system" field
* tools: parameters → input_schema, no "type:function" wrapper
* tool_choice "auto" → {"type":"auto"} object
* assistant tool calls → content[{type:tool_use, id, name, input}]
* role=tool results → role=user content[{type:tool_result,...}]
* consecutive tool results merged into single user turn
- convertMessagesToAnthropic: handles all three message types
- parseMCPResponseFull: extracts text + tool_use blocks
- parseMCPResponse: delegates to parseMCPResponseFull
All mcp and agent tests pass.
* fix(telegram): fix claude client dispatch + strategy creation workflow
- telegram/bot.go: clientForProvider now returns NewClaudeClient() for
'claude' provider (was incorrectly falling back to DeepSeekClient which
uses OpenAI wire format, breaking Anthropic API calls)
- api/server.go: fix scan_interval_minutes schema default (3, not 60);
POST /api/strategies now clearly states config is OPTIONAL with complete
working defaults; POST /api/traders removes redundant GET workflow note
- telegram/agent/prompt.go: simplify strategy creation — just POST {name}
without config (backend applies full working defaults automatically);
only include config when user requests custom settings
* test(mcp): add ClaudeClient wire format tests
Tests cover all Anthropic-specific format conversions:
- system prompt lifted to top-level field
- tools use input_schema (not parameters)
- tool_choice is object {type:auto} not string
- assistant tool calls → content[{type:tool_use}]
- consecutive tool results merged into single user turn
- parseMCPResponseFull: text, tool_use, and error cases
- x-api-key header (not Authorization: Bearer)
- /messages endpoint URL
* fix(telegram): clientForProvider returns correct client for all 7 providers
Previously qwen/kimi/grok/gemini all fell back to DeepSeekClient.
Each provider now gets its own dedicated client with correct default
base URL and model. All 7 providers now fully supported:
openai, deepseek, claude, qwen, kimi, grok, gemini
* fix(telegram): newLLMClient uses bound user's model, not any user's model
GetAnyEnabled() searched across all users in DB — if user B has an
enabled model, bot could use their API key while acting as user A.
Now uses GetDefault(botUserID) which only looks up the bound user's
enabled model, matching the same user scope as all API calls.
* fix(auth): single-user deployment by default, no open registration
Registration logic redesigned:
- Empty DB (first-time setup): registration always open, no config needed
- After first user exists: registration closed by default
- Multi-user opt-in: set REGISTRATION_ENABLED=true + MAX_USERS=N in .env
Config defaults changed:
- RegistrationEnabled: true → false (closed after first user)
- MaxUsers: 10 → 1 (single-user deployment default)
This eliminates the confusion of multiple users appearing in a personal
deployment where Telegram is bound to a single admin account.
* feat(solo): beginner-friendly onboarding — smart setup guide + direct config commands
start.sh:
- Interactive Telegram Bot Token prompt on first run
- Token format validation (must match 12345:ABC... pattern)
- Friendly step-by-step startup instructions after launch
telegram/bot.go:
- /start now shows context-aware setup guide based on actual config state:
- No AI model → explains how to configure, lists all providers
- AI model OK but no exchange → guides to configure exchange via chat
- All configured → full capabilities welcome message
- New: direct setup commands ('配置 deepseek sk-xxx') bypass LLM entirely
so AI model can be configured even before any model exists (bootstrap fix)
- All messages now in Chinese (匹配用户语言)
telegram/agent/prompt.go:
- Added first-time setup detection section
- Agent told to never ask user to visit web UI — everything via chat
* feat(i18n): bilingual EN/ZH setup guide with language selection
store/telegram_config.go:
- Add Language field to TelegramConfig (persisted in DB)
- Add SetLanguage(lang) and GetLanguage() methods
- Default language: English (en)
telegram/bot.go:
- First /start triggers language selection (1=English, 2=中文)
- /lang command to change language at any time
- awaitingLang state machine handles language choice before any other input
- buildSetupGuide() now fully bilingual (EN/ZH), context-aware:
Step 1: configure AI model (no model yet)
Step 2: configure exchange (model OK, no exchange)
Ready: show full capabilities
- tryHandleSetupCommand() bilingual: 'configure/配置 <provider> <key>'
- helpMessage(lang) fully bilingual
- All error/status messages bilingual
Default: English. isLangDefault() detects whether user has explicitly
chosen a language vs falling back to the 'en' default.
* fix(telegram): use Markdown rendering + simplify language selection condition
- sendMarkdownMsg() helper: sends with ParseMode=Markdown, falls back to plain text
- All formatted messages (langSelectionMsg, buildSetupGuide, helpMessage) now render
bold text and code blocks correctly in Telegram
- Simplify /start language check: isLangDefault(st) alone is sufficient
(lang == 'en' && isLangDefault was redundant — GetLanguage returns 'en' when empty)
* fix(start.sh): translate all user-facing text to English
Entire script was in Chinese. Now English-first throughout:
- startup banner, prompts, success/error messages
- setup_telegram(): English instructions and validation messages
- start(): English next-steps after launch
- stop/restart/clean/update/regenerate-keys/show_help: all English
* fix(telegram): remove 'default' user fallback — resolve user dynamically
- botUserID no longer captured once at startup (was 'default' if no user yet)
- resolveBotUser() reads first registered user from DB on demand:
* called on every /start (handles: registered after bot launch)
* called before every AI message (handles mid-session registration)
- If no user registered: clear English error 'No account found. Please register on the web UI first'
- start.sh: fix set_env_var appending without newline (token was concatenated to prev line)
* refactor(telegram): clean onboarding — web UI for setup, Telegram for operations
- /start shows clean status: 'setup required → open web UI' or 'ready → examples'
- Removed tryHandleSetupCommand (no more CLI-style 'configure deepseek sk-xxx')
- Removed automatic language selection on /start (use /lang anytime instead)
- newLLMClient returns nil when no model → clear guard, not fallback
- statusMsg() replaces buildSetupGuide(): two states only (missing config / ready)
- Bot is now purely an operations interface; config lives in the web UI
* refactor: single-user web-based setup — replace env config with Settings UI
Move from multi-user env-var config to single-user web-first architecture:
- Add SetupPage for first-time initialization (replaces /register)
- Add SettingsPage for AI models, exchanges, Telegram, and password management
- Enrich all API route schemas with exact ID usage documentation
- Add PUT /user/password endpoint for in-app password changes
- Remove REGISTRATION_ENABLED, MAX_USERS, TELEGRAM_BOT_TOKEN from env config
- Simplify LoginPage design, remove admin mode and registration links
- Telegram bot now resolves user email for identity display
- start.sh no longer runs interactive Telegram setup
* feat: add blockRun (x402 USDC) support to all AI model consumers
- telegram/bot.go: add blockrun-base, blockrun-sol, minimax to
clientForProvider; fix newLLMClient to prefer TelegramConfig.ModelID
over GetDefault; log USDC payment provider usage
- debate/engine.go: add blockrun-base, blockrun-sol to InitializeClients
- api/strategy.go: add blockrun-base, blockrun-sol to runRealAITest
- backtest/ai_client.go: add blockrun-base, blockrun-sol to configureMCPClient
* feat: add Claw402 (claw402.ai) x402 USDC payment provider
Add Claw402Client for claw402.ai's x402 micropayment gateway (Base USDC).
Supports 15+ AI models (GPT-5.4, Claude Opus, DeepSeek, Qwen, Grok, etc.)
with per-model endpoint routing.
- mcp/claw402.go: new client with model→endpoint mapping, x402 v2 payment flow
- mcp/blockrun_base.go: extract shared signX402Payment() for reuse
- Register "claw402" provider in all 6 consumer switch statements:
api/server.go, api/strategy.go, trader/auto_trader.go,
telegram/bot.go, debate/engine.go, backtest/ai_client.go
* feat: redesign Claw402 model config UI — friendly wallet setup, USDC guide, official logo, nginx no-cache for index.html
* refactor: centralize x402 payment flow into shared mcp/x402.go
Extract duplicated doRequestWithPayment/call/CallWithRequestFull/buildRequest/
setAuthHeader (~165 lines x3) into shared helpers in mcp/x402.go. Consolidate
shared types (x402v2PaymentRequired, x402AcceptOption, x402Resource) and remove
duplicate Solana types. Fix validAfter to 0 (official SDK standard), drain 402
body before retry, log Payment-Response tx hash, check Payment-Required before
X-Payment-Required.
* fix: stop PR template bot from overwriting user-written descriptions
The pr-template-suggester workflow was triggered on opened/edited/synchronize
events and forcefully replaced the PR body with a template when body < 100 chars.
This caused user-written descriptions to be overwritten.
Replace with a lightweight labeler (OpenClaw-style) that:
- Only adds labels (backend/frontend/docs, size: XS/S/M/L/XL)
- Never modifies the PR body
- Simplified unified PR template at .github/pull_request_template.md
* chore: simplify PR template (OpenClaw-style)
Integrates BlockRun (blockrun.ai) as a new AI provider option via x402
micropayment protocol, allowing users to access top AI models with USDC
without requiring individual API keys.
- Add BlockRun Base (EVM) and Solana wallet providers to model selector
- Implement x402 v2 EIP-712 payment signing for Base (mcp/blockrun_base.go)
- Implement x402 v2 SPL TransferChecked signing for Solana (mcp/blockrun_sol.go)
- Wire blockrun-base and blockrun-sol into trader factory (auto_trader.go)
- Register both providers in supported models API (server.go)
- Add BlockRun card UI with wallet key input in Step 0/1 of model config modal
- Add BlockRun SVG icon and ModelIcons support
- Add setup guides for Base and Solana wallet configuration (docs/)
- Available flagship models: GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro,
Grok 3, DeepSeek Chat, MiniMax M2.5
