Skip to content
CAI
Software that uses CAICheck a score

HKUDS/Vibe-Trading

56.9

Weak · 16 November 2026

292.8k

lines of production code

Python

with TypeScript

5

measurements over time

CAI band scale
CAI trend line
CAI lens gauges

What this system is

Features

Add Korea Investment & Securities (KIS) trading connector

Introduces a new broker connector for Korea Investment & Securities (KIS), enabling access to Korean equities (KOSPI/KOSDAQ) via the official KIS Developers REST API. The connector supports paper trading (both read-only and order placement) and live read-only access, leveraging a structural separation between paper and live hosts to prevent accidental live orders from paper profiles. It includes built-in profiles, operation classification for read/write gating, and handles OAuth2 token management, KST time zones, and multi-venue order history retrieval (KRX, SOR, NXT).

agent/src/trading/connectors/kis · high confidence

Add Mootdx-based A-share data fetching capability

The agent/skills area now includes a new A-share data fetcher (a\_mootdx\_fetcher) that retrieves daily and real-time market data for Chinese stocks using the Mootdx library. This adapter supports fetching historical bars and current quotes without requiring an API key or incurring IP rate limits, and includes logic to handle batch requests with appropriate delays to respect the provider's constraints.

agent/skills · high confidence

Add Shoonya (Finvasia) trading connector for Indian markets

Introduces a new connector for Shoonya (Finvasia), providing read-only and paper-trading access to NSE/BSE equities, F&O, currency, and commodity markets via the NorenRestApiPy SDK. The implementation includes configuration profiles for paper and live-read-only modes, a tool classification map to enforce read/write permissions, and SDK wrappers for market data and order simulation. Live order placement is structurally disabled as the broker API lacks a runtime paper/live discriminator.

agent/src/trading/connectors/shoonya · high confidence

Add Toss Securities (토스증권) read-only trading connector

Introduces a new read-only connector for Toss Securities, enabling access to account snapshots, positions, open orders, quotes, and historical candle data via the official Toss Invest Open API. The implementation includes configuration management, tool classification (read/write gating), and profile definitions. Due to the absence of a verifiable sandbox environment, order placement and cancellation are strictly disabled for all profiles, ensuring the connector operates in a safe, read-only mode for both live and paper trading contexts.

agent/src/trading/connectors/toss · high confidence

Add Upbit broker connector for Korean crypto markets

Introduces a new trading connector for Upbit, Korea's largest KRW crypto exchange, enabling read-only access to live account data and locally simulated paper trading. The implementation uses Upbit's public REST API with JWT authentication, requiring no third-party SDKs. It provides three built-in profiles: a read-only live profile, a read-only paper profile for market data, and a paper-trade profile that simulates order placement locally against live ticker prices. The connector is structurally capped at paper for order execution because Upbit offers no sandbox or runtime paper/live discriminator, ensuring no real funds are at risk during simulation.

agent/src/trading/connectors/upbit · high confidence

Add Zerodha Kite Connect trading connector

Introduces a new connector for the Zerodha Kite Connect API, enabling access to Indian equity and F&O markets (NSE/BSE). The implementation provides read-only capabilities for historical OHLCV data, quotes, positions, and holdings, alongside a locally simulated paper-trading mode for order placement. Live order placement is explicitly disabled to prevent accidental execution against real accounts, as the API lacks a runtime paper/live discriminator. The connector handles Kite-specific constraints, such as paginating historical data requests within API limits and rejecting unsupported intervals like 4h.

agent/src/trading/connectors/zerodha · high confidence

Add ZimaOS deployment template and documentation

Introduces a new \zimaos/\ directory containing a Docker Compose template, a README with configuration and usage instructions, and an example GitHub Actions workflow for fork-based image publishing. The Compose setup runs the Vibe-Trading application on port 8899 with persistent volumes for runs, sessions, and Brazilian market data, while explicitly disabling automatic updates via Watchtower by default. It also documents the shift of the Brazilian data sync from a sidecar container to a protected system job within the main service, requiring explicit environment variables to enable.

zimaos · high confidence

The wiki now displays aggregate traffic statistics in the footer, distinguishing between human visitors, AI agents (such as GPTBot and ClaudeBot), and generic bots. This data is collected server-side via a new middleware that writes daily counts to a D1 database without storing personal identifiers, and exposed through a new /api/stats endpoint that also caches PyPI install counts for the vibe-trading-ai package to ensure reliability.

wiki/functions · high confidence

Add qlib158 factor zoo with 154 alpha features

The qlib158 factor zoo is now available, providing 154 alpha features adapted from Microsoft's Qlib project. This includes K-bar features (e.g., KMID, KLEN), rolling features (e.g., ROC, MA, STD, BETA, CORR, CORD, CNTD, CNTP, CNTN, WVMA) across 5, 10, 20, 30, and 60-day windows, and momentum indicators (e.g., IMAX). The implementation is a clean-room re-expression using the project's base operators, ensuring no direct code copying from the upstream Apache-2.0 licensed source. Factors are designed for daily frequency data across US, CN, HK, IN, and KR equity universes.

agent/src/factors/zoo/qlib158 · high confidence

Centralized, typed configuration layer with safe environment parsing and atomic config writes

The agent now uses a unified configuration system that replaces scattered environment variable reads with a validated, Pydantic-based schema. This ensures that optional numeric environment variables fall back to defaults instead of crashing initialization, and provides a single source of truth for all agent settings. The system includes a thread-safe singleton accessor for runtime configuration, backward-compatible alias support for renamed environment variables, and atomic file writing for channel credentials to prevent data corruption. Additionally, it introduces a migration path to move legacy session and run data from the code directory to a dedicated runtime root, ensuring history is preserved across upgrades.

agent/src/config · high confidence

Expanded language support with lazy loading and RTL layout handling

The application now supports nine languages: English, Chinese (zh-CN), Japanese, Korean, Arabic, Spanish, German, Brazilian Portuguese, and Indonesian. This change introduces a lazy-loading mechanism for non-English locales to improve initial load performance and adds support for right-to-left (RTL) layouts, specifically for Arabic, by automatically adjusting the document direction. The language preference is persisted in local storage, and the system gracefully handles loading failures by reverting to the previous stable language.

frontend/src/i18n · high confidence

Introduce Alpaca trading connector with paper/live separation and mandate-gated live trading

Adds a new Alpaca connector (agent/src/trading/connectors/alpaca) that exposes read-only access for both paper and live accounts, plus order placement capabilities. Paper and live environments are structurally separated using distinct API keys and hosts (paper-api.alpaca.markets vs api.alpaca.markets), ensuring paper keys cannot reach the live host. Live order placement is gated by a mandate requirement (orders.place.requires\_mandate), meaning users must have an authorized mandate before placing orders on live funds. The connector supports both direct SDK usage via alpaca-py and an opt-in TAP routing mode for credential isolation, where broker egress is routed through a TAP proxy with human approval for writes. Configuration is managed via alpaca.json with support for paper, live-readonly, and live profiles, including feed selection (iex/sip).

agent/src/trading/connectors/alpaca · high confidence

Introduce Alpha Zoo framework with registry, bench runner, and comparison tools

The \agent/src/factors\ module now provides the core infrastructure for the Alpha Zoo, a framework for managing and evaluating 460 alphas across five categories (alpha101, gtja191, qlib158, academic, fundamental). This change introduces a \Registry\ for alpha management, a \bench\_runner\ for computing Information Coefficient (IC) statistics and categorizing alphas as 'alive', 'reversed', or 'dead', and a \compare\_runner\ for head-to-head alpha comparisons. It also adds CLI handlers (\cli\_handlers.py\) for listing, showing, benchmarking, and comparing alphas, along with a \factor\_analysis\_core\ for computing IC series and layered backtest equity curves. The backend (\\_backend.py\) implements lazy initialization of the \bottleneck\ library for performance optimization, with a graceful fallback to pandas if unavailable.

agent/src/factors · high confidence

Introduce Binance trading connector with read-only and mandate-gated order placement

This change adds the Binance trading connector module, providing structured access to Binance markets via the ccxt library. It introduces four distinct profiles: read-only access for both paper (testnet) and live environments, plus order-placement capabilities for paper (testnet) and live (mainnet). The live order-placement profile is strictly gated behind a user mandate requirement. Additionally, the connector supports a read-only USD-M (USDT-Margined) account mode for live environments, which performs strict validation of account and position data against Binance's USD-M endpoints to ensure data integrity and security.

agent/src/trading/connectors/binance · high confidence

Introduce CLI for managing Pixel Office agents and tasks

A new command-line interface has been added to the \agent/src/pixel\_agents\ module, allowing users to initialize an office database, register agents with specific roles and specialties, create and assign tasks, complete tasks with results, adjust agent ranks, and view a dashboard or list of current agents and tasks. This tool provides a manual research planning interface that manages office state via SQLite, explicitly excluding task dispatch or trading authority.

_agent/src/pixel\agents · high confidence

Introduce Dhan broker connector for Indian markets

Added a new trading connector for the Dhan broker, supporting NSE/BSE equities and F&O markets via the official dhanhq SDK. The implementation includes read-only access for both paper and live environments, and a paper-only mode for order placement that simulates trades locally using real market data. This addresses the lack of a runtime sandbox by structurally capping live profiles to read-only operations while allowing safe paper trading.

agent/src/trading/connectors/dhan · high confidence

Introduce Futu (moomoo) trading connector with read-only and mandate-gated live profiles

Added a new connector for Futu (moomoo) that connects to a local OpenD gateway (default 127.0.0.1:11111) via the official futu-api SDK. The connector exposes read-only capabilities for account, positions, orders, quotes, and historical data, along with extended read endpoints like capital flow and earnings calendars. It enforces strict paper/live separation by resolving accounts based on the trd\_env field (SIMULATE vs REAL) and fails closed if there is a mismatch. Live order placement is supported in a dedicated profile but is gated behind a user mandate and requires unlocking the trade context with a hashed password, ensuring no live orders can be placed without explicit authorization.

agent/src/trading/connectors/futu · high confidence

Introduce Longbridge trading connector with read-only and paper-trade profiles

Added a new Longbridge (LongPort OpenAPI) trading connector that wraps the official Python SDK for account and market data access. The connector ships with built-in profiles for read-only live access and paper trading, while explicitly restricting live order placement due to the SDK's lack of a runtime discriminator between paper and live environments. It includes a tool classification map to enforce read/write gating and handles configuration via a dedicated JSON file or API key resolution.

agent/src/trading/connectors/longbridge · high confidence

Introduce OKX trading connector with paper/live profiles and mandate-gated live trading

Added the OKX trading connector, providing read-only access to account balances, positions, orders, fills, tickers, and candlesticks via the python-okx SDK. The connector supports distinct paper (demo) and live environments, selected via SDK flags and separate key namespaces. It includes built-in profiles for read-only paper/live access and trading capabilities, with live order placement (place, cancel, amend) gated behind a user-defined mandate to ensure safety.

agent/src/trading/connectors/okx · high confidence

Introduce Robinhood connector with mandate enforcement and read-only portfolio support

The Robinhood connector now includes a curated read/write classification map and an order-intent extractor that validates buy/sell orders against mandate requirements, ensuring that order sizing is enforced at the worse of the quote and limit price. A new read-only portfolio source provides account-bound live trading data without execution capabilities, while the connector also aligns with the MCP startup process and forwards account numbers to remote calls for proper scoping.

agent/src/trading/connectors/robinhood · high confidence

Introduce Shadow Account for strategy extraction and backtesting

Users can now extract trading patterns from their broker journals into a 'Shadow Profile' and backtest them across multiple markets (China A-share, HK, US, Crypto). The system identifies profitable roundtrips, distills them into decision rules, and runs a multi-market backtest that splits runs by settlement currency to avoid FX issues. It provides delta-PnL attribution (noise, missed signals, overtrading) and generates an 8-section report with charts, available as HTML, PDF (via weasyprint or a reportlab fallback), or HTML-only if rendering engines are missing.

_agent/src/shadow\account · high confidence

Introduce Strategy Development Manager artifact store and decay monitoring

Added a new strategy store subsystem under \agent/src/strategy\_store\ that provides a persistent, SQLite-backed (with an in-memory reference implementation) registry for strategy and factor artifacts. This change introduces core capabilities for the Strategy Development Manager, including artifact lifecycle management (creation, status transitions, and governance metadata like validation status and model tier), bench history recording with deduplication, and automated decay monitoring. The store computes rolling and baseline metrics (IC, Sharpe, IR) from bench history to detect performance decay, ensuring that non-finite values are excluded and that metric calculations handle edge cases like negative baselines correctly.

_agent/src/strategy\store · high confidence

Introduce Tiger Brokers trading connector with paper and live profiles

Added a new connector for Tiger Brokers (TigerOpen) that enables read-only access to account, market, and order data, as well as order placement capabilities. The implementation includes four built-in profiles: paper and live read-only modes for safe data inspection, a paper trading mode for sandbox order execution, and a live trading mode that enforces a mandate gate (including symbol allowlists and exposure limits) before placing orders. The connector uses RSA-signed REST authentication via the official Python SDK and validates account types to prevent misconfiguration between paper and live environments.

agent/src/trading/connectors/tiger · high confidence

Introduce Vibe-Trading Desktop (Unofficial Community Build) for Windows

This change adds a new Electron-based desktop host for the Vibe-Trading application, targeting Windows packaging and lifecycle management. The desktop shell enforces a single application instance, starts the backend on a random loopback port, and generates a per-process authentication secret to secure API requests without exposing them to the renderer. It includes a parent-death watchdog to ensure the backend process is terminated if the Electron main process dies unexpectedly, and implements secure credential storage using Electron's safeStorage, migrating existing plaintext secrets from environment files. The package also provides an unsigned NSIS installer artifact for review, a checksum-pinned Python 3.12.10 runtime, and localized UI strings for English, Simplified Chinese, Japanese, Korean, and Arabic. Dormant safety primitives for a future signed updater are included but not enabled. Also in this area: 5 commits (2 fixes) modifying desktop.

desktop/electron · medium confidence · unverified

Introduce Vibe-Trading Wiki home page with bilingual support and installation guides

The Vibe-Trading Wiki now includes a dedicated home page that introduces the platform as an open-source finance research agent. This page provides users with clear installation instructions via PyPI, uv, or source, and outlines the core research workflow (Route, Ground, Test, Deliver) and key capabilities like Agent Harness, Backtesting, Swarm Teams, and Shadow Account. A key addition is the built-in language toggle, allowing users to switch the interface between English and Chinese, making the documentation more accessible to a broader audience.

wiki/home · high confidence

Introduce WebSocket channel runtime and compatibility helpers

Added a new \channelsui\ module providing the runtime infrastructure for the WebSocket channel, including gateway services for workspace scoping and token issuance, HTTP utilities for route normalization and file uploads, and API helpers for normalizing CLI app and MCP preset mentions. The change also adds media decoding utilities to handle base64 data URLs and introduces compatibility shims that return structured errors for currently unsupported features like fork-chat and audio transcription.

agent/src/channelsui, agent/src/utils · high confidence

Introduce bounded-autonomy live trading channel with mandate enforcement and kill switch

The agent now supports a new live trading channel that allows autonomous order execution within a user-committed mandate and ring-fenced funds. This change introduces a pre-trade enforcement gate (LiveOrderGuardTool) that validates mandates, checks expiry, enforces a filesystem-based global and per-broker kill switch, and applies daily order limits. It also adds a 3-tier tool classification system (annotations, curated maps, default-deny) to safely wrap broker tools, ensuring write operations are always gated. The implementation includes direct-SDK support for brokers like Alpaca and Tiger, alongside MCP-based wrappers for others, with full audit trails and cross-platform file locking for order counters.

agent/src/live · high confidence

Introduce bounded-autonomy mandate system for live trading

Added a new mandate subsystem in \agent/src/live/mandate\ that enforces user-defined safety boundaries (hard caps, universe constraints, and consent provenance) for live trading. The system includes a read-only loader (\store.py\) that parses mandate files at boot and a commit module (\commit.py\) that handles the secure, atomic persistence of mandate profiles and consent records via a dedicated API endpoint, ensuring the agent loop cannot self-authorize changes.

agent/src/live/mandate · high confidence

Introduce cross-market composite backtest engine with per-market rule sets

The backtest engine now supports running a single simulation across multiple asset classes (equities, futures, crypto, forex) using a shared capital pool. A new CompositeEngine routes symbols to specialized sub-engines—such as ChinaAEngine, ChinaFuturesEngine, BrazilEquityEngine, and GlobalEquityEngine—each enforcing its own market-specific rules (e.g., T+1 settlement, stamp duties, lot sizes, and price limits). Symbol classification is centralized in \_market\_hooks to ensure consistent routing, and the system prevents mixed-currency runs to maintain the integrity of the shared equity curve.

agent/backtest/engines · high confidence

Introduce eToro Public API connector with demo and live profiles

Adds a new eToro connector that supports both demo (paper) and live trading via the eToro Public API, using a single API key pair where the profile selects the API path. The implementation includes read operations (status, profile, positions, orders, quotes, instrument search and metadata), write operations (place, cancel, close positions, edit stops), and copy-trading workflows (precheck, start/adjust, poll, close). It enforces structural safety by blocking copy trading on demo accounts and requiring a mandate for live write actions, while also handling instrument type resolution and PnL reading from the aggregate portfolio route.

agent/src/trading/connectors/etoro · high confidence

Introduce entity and irregular cash-flow spine for non-bar holdings

A new \agent/src/entities\ package provides a parallel path to the existing backtest engine, specifically designed for holdings that do not have daily price bars. This module introduces \CashFlowSeries\ and related models (such as \Entity\, \Fund\, and \Bond\) to handle irregular cash flows and NAV-based analytics, ensuring that non-OHLCV data is processed by cash-flow-native logic rather than being incorrectly priced as standard close prices by the bar engine.

agent/src/entities · high confidence

Introduce evidence-gated strategy discovery with per-regime backtest validation

The \agent/src/strategy\_discovery\ package now provides a unified, read-only catalog of strategies from Alpha Zoo and the SDM store, where all performance metrics are strictly gated on real backtest evidence rather than seed data. The system computes per-regime (bull, bear, structural) metrics from reproducible run artifacts, enforcing hard quality gates such as minimum trade counts, coverage days, and evidence freshness to prevent stale or insufficient data from influencing recommendations. It also introduces a description-driven adaptation workflow that allows creating child strategies from parent evidence, complete with integrity checks to ensure adaptation artifacts remain sealed and bound to their specific run directories.

_agent/src/strategy\discovery · high confidence

Introduce finance research goal runtime with lifecycle and policy enforcement

The agent now supports a dedicated finance research goal runtime, introducing a structured lifecycle (active, paused, complete, etc.) and a criterion-based evidence ledger to track research progress. This includes new data models for goals, claims, and evidence, along with context formatting helpers that inject goal state and open criteria into the agent's prompt to guide multi-turn research. Policy enforcement is added to reject objectives that resemble live trading or order execution, ensuring the research scope remains compliant.

agent/src/goal · high confidence

Introduce governance primitives for run manifests and audit ledgers

The agent now exposes a governance module providing deterministic run manifests and a tamper-evident hash-chained ledger. Users can generate a manifest hash that fingerprints a run's methodology (system prompt, skills, tools, and package versions) and compare manifests to identify specific changes between runs. Additionally, the module provides an append-only, fsynced JSONL ledger with chain verification capabilities, which can be integrated into live action logging via an opt-in mode.

agent/src/governance · high confidence

Introduce hypothesis registry and CLI management commands

Users can now manage research hypotheses via a new durable registry and CLI commands. The \vibe-trading hypothesis list\ command displays hypotheses with optional status filtering and JSON output, while \show\ provides detailed views including thesis, signals, and linked run cards. The \invalidate\ command allows marking hypotheses as rejected with optional notes. Storage is handled via a configurable path (defaulting to a standard location, overridable via \VIBE\_TRADING\_HYPOTHESES\_PATH\ env var or \--path\ flag), supporting statuses like exploring, testing, validated, rejected, and monitoring.

agent/src/hypotheses · high confidence

Introduce local and official MCP read-only Interactive Brokers connectors

Users can now connect to Interactive Brokers via two new read-only modes: a local connector that communicates with a user-owned TWS or IB Gateway instance (supporting both paper and live accounts with configurable market-data tiers), and an official MCP connector that uses IBKR's OAuth-approved read-only tools for account summaries and positions. The implementation includes a thread-safe connection pool, snapshot-based quote retrieval, and strict classification of MCP tools to ensure fail-closed security for state-changing operations.

agent/src/trading/connectors/ibkr · high confidence

Introduce persistent cross-session memory with compression, lifecycle management, and semantic linking

Added a new persistent memory subsystem in agent/src/memory that stores cross-session data as files. The system includes a three-level compression pipeline (raw, daily, digest) using TF-IDF to extract key sentences, a lifecycle manager for quality scoring, importance decay, and garbage collection (archiving low-importance entries), and a semantic linker that creates BM25-based relationships between memories. These capabilities are gated behind feature flags (VT\_MEMORY\_COMPRESSION, VT\_MEMORY\_QUALITY, VT\_MEMORY\_GC, VT\_MEMORY\_DECAY, VT\_MEMORY\_LINKS) and can be triggered via a scheduled consolidation job.

agent/src/memory · high confidence

Introduce pluggable OCR engine architecture with local and cloud backends

The OCR tooling now supports a pluggable engine architecture, allowing users to choose between a local RapidOCR engine (privacy-preserving, no API key required) and a cloud-based LLM vision engine (via any OpenAI-compatible provider). Engine selection is controlled by the VIBE\_TRADING\_OCR\_ENGINE environment variable, which accepts 'auto' (default, selects the first available local engine), a specific engine name, or 'none' to disable OCR. The system automatically discovers third-party engines installed via pip entry points and includes backward-compatibility aliases for deprecated engine names. This change provides greater flexibility and privacy control over document text extraction.

agent/src/tools/ocr · high confidence

Introduce read-only multi-broker portfolio aggregation and dashboard

Users can now view a unified, read-only portfolio dashboard that aggregates holdings from multiple broker connections and local CSV imports. The new \agent/src/portfolio\ module provides a \PortfolioService\ that normalizes data from various brokers (including specific handling for Futu, IBKR, Longbridge, Binance, and OKX), validates display currencies against available FX rates, and stores immutable snapshots in a local SQLite database. This feature supports configuring multiple read-only sources, importing extraETF holdings, and generating deterministic allocation proposals without executing trades.

agent/src/portfolio · high confidence

Introduce scheduled research job persistence and execution

Users can now create, manage, and run scheduled research jobs that persist across restarts and execute automatically on a defined schedule. This change adds a crash-safe JSON store for job state, an executor that polls for due jobs and dispatches them (including non-blocking fire-and-forget execution for internal jobs), and a proposal system to safely create or cancel jobs while protecting system jobs from user mutation. Jobs support timezone-aware cron schedules, configurable delivery channels with PDF password protection, and verdict recording for each run.

_agent/src/scheduled\research · high confidence

Introduce shared scheduling kernel and internal callable registry

The agent now uses a unified scheduling infrastructure in \agent/src/scheduling\. A new \SchedulerKernel\ provides a generic background tick loop with configurable retry backoff, an injectable clock for testing, and an optional adaptive sleep hint to reduce polling overhead. An explicit allowlist registry (\src.scheduling.registry\) restricts internal scheduled jobs to pre-registered callables, preventing arbitrary code execution. The registry currently supports two internal jobs: \memory.consolidate\ for scheduled memory consolidation and \br\_data.run\_sync\ for the Brazilian official-dataset sync. Shared trigger grammar (interval-ms and 5-field cron) and type definitions are centralized to ensure consistent scheduling behavior across the application.

agent/src/scheduling · high confidence

Introduce the quantlib package with auditable financial-math primitives

The new agent/src/quantlib package replaces the previous practice of carrying financial-math formulas as markdown in skills (which led to unreviewable, non-reproducible code) with a single, tested, and importable Python implementation. This package provides a comprehensive set of modules for options pricing (Black-Scholes, barrier options, implied volatility), risk analytics (VaR, CVaR, drawdown, Monte Carlo), portfolio allocation (Hierarchical Risk Parity), market microstructure signals (VPIN, Roll spread, Amihud, Kyle), copula models (Archimedean, Gaussian), and valuation models (DCF, comps, three-statement). It also includes time-series tests (ADF, cointegration, GARCH) and market impact models. The package is designed to be robust, with strict input validation, handling of degenerate cases, and lazy loading of optional dependencies like statsmodels and arch to ensure the base install remains lightweight.

agent/src/quantlib · high confidence

Kakushadze 101 Formulaic Alphas factor zoo added

The \agent/src/factors/zoo/alpha101\ directory now contains 101 standalone Python modules implementing the Kakushadze (2015) formulaic alphas. Each module provides a \compute\ function that calculates its specific alpha signal on OHLCV+ market data panels, with metadata in \\_\_alpha\meta\\_\ defining required columns, universes (US, India, Korea), and warmup periods. The implementation includes a LICENSE.md documenting the source paper and a CI gate to prevent trademark strings, ensuring the factors are ready for use in backtesting and live trading strategies.

agent/src/factors/zoo/alpha101 · high confidence

Launch of static Vibe-Trading wiki with bilingual support and traffic analytics

The wiki section is now available as a static site at vibetrading.wiki, featuring a unified English and Chinese interface with a persistent language toggle and a dark/light theme switcher. The site includes dedicated sections for documentation, tutorials, the Alpha Library, and the research lab, with routing handled via Cloudflare Pages. Additionally, an anonymous, first-party traffic counter is implemented to display aggregate visits from AI agents and humans in the footer, backed by a Cloudflare D1 database.

wiki · high confidence

New Brazilian market data sync module with SQLite storage and scheduler integration

A new \agent/src/br\_data\ module has been added to sync official Brazilian market data (B3 COTAHIST, CVM funds/companies/statements, Banco Central series, and Tesouro Direto) into a local SQLite database. The module provides a CLI (\python -m src.br\_data sync\) for manual updates and integrates with the main service's scheduler via a protected system job (\system:br-data-sync\) that runs daily. The sync is opt-in, controlled by the \VIBE\_BR\_SYNC\_ENABLED\ environment variable, and defaults to a 21:30 São Paulo time slot (configurable via \VIBE\_BR\_SYNC\_TIME\). The standalone daemon command is retired in favor of this scheduler-based approach.

_agent/src/br\data · high confidence

New CLI formatting and status helpers

The CLI now uses a new set of pure utility functions for displaying status information. Duration, token counts, and numeric values are formatted consistently with the web app (e.g., durations show sub-second precision like "1.4s", and token counts use suffixes like "1.2k"). Additionally, the streaming spinner now displays randomized "thinking" verbs (such as "Pondering…" or "Analyzing…") to provide more dynamic feedback during agent generation.

agent/cli/utils · high confidence

New LLM usage and positions visualization panels in Run Detail

The Run Detail view now includes dedicated panels for LLM token usage and positions structure. The LlmUsagePanel displays a summary of provider, model, and token counts (including optional cache read/creation tokens) alongside a per-iteration breakdown table. The PositionsTab visualizes portfolio holdings using pie charts or treemaps, supports date-based filtering, and classifies assets by market (e.g., A-shares, US equity) with sector aggregation. These components rely on shared RunCard building blocks (RunCardPanel, KeyValueTable) extracted to avoid circular imports.

frontend/src/components/run · high confidence

New MetaTrader 5 connector for Exness-style brokers

Users can now connect to MetaTrader 5 terminals (specifically Exness-style brokers) via a new optional connector. This feature adds support for reading account status, positions, and market data, as well as placing and managing orders on both demo and live accounts. The connector requires the Windows-only \MetaTrader5\ Python package and a locally running terminal, with configuration managed through \\~/.vibe-trading/mt5.json\. It includes built-in safeguards to prevent mixing paper and live profiles and enforces mandate-based controls for live trading.

agent/src/trading/connectors/mt5 · high confidence

New Shadow Account HTML report and generated signal engine

The Shadow Account feature now includes a styled HTML report (shadow\_report.html) and its accompanying CSS (shadow\_report.css) that visualize the comparison between real trading performance and the shadow backtest, including delta PnL, rule coverage, and attribution of performance differences. Additionally, a Jinja2 template (signal\_engine.py.j2) generates a Python signal engine that applies the extracted shadow rules to market data, correctly handling market-specific rule matching and entry conditions like RSI and prior returns while skipping time-of-day gates for daily bars.

_agent/src/shadow\account/templates · high confidence

New Windows packaging, security, and lifecycle test infrastructure

This change introduces a suite of new scripts in the desktop build pipeline to harden Windows packaging and verify runtime behavior. The \build-backend.ps1\ script now handles the secure, checksum-verified download and extraction of the Python 3.12.10 runtime and GTK dependencies, while \build-review-installer.mjs\ and \build-signed-installer.mjs\ provide separate paths for unsigned review builds and Authenticode-signed production releases. To ensure reliability, \prepare-electron.mjs\ manages the Electron runtime with bounded retries and SHA-256 verification, and \process-utils.ps1\ provides a robust process execution wrapper. The update flow is secured by \test-update-safety.mjs\, which validates update artifacts against publisher certificates and version policies, and \smoke-lifecycle.mjs\ and \smoke-parent-death.mjs\ verify that the backend lifecycle and parent-process death cleanup operate correctly without affecting unrelated processes. Also in this area: 5 commits (2 fixes) modifying desktop.

desktop/electron/scripts · medium confidence · unverified

New academic factor zoo with price-based proxies and missing-data safeguards

The academic factor zoo now includes a suite of canonical academic factors (MKT-RF, SMB, HML, RMW, CMA, Carhart UMD, BAB, 52-week-high, Amihud illiquidity, short-term reversal, return skewness, and correlation-rewiring stability) implemented as price/volume proxies using only OHLCV inputs. Each factor is cross-sectionally z-scored for long-short ranking, with explicit disclosures that they approximate original fundamental-based definitions. The implementation enforces strict missing-data handling: returns use fill\_method=None to prevent silent forward-filling, short panels produce all-NaN results surfaced as RegistryErrors rather than misleading shrunk-window values, and the correlation-rewiring factor requires minimum coverage thresholds to keep missing bars missing under pandas 2.x.

agent/src/factors/zoo/academic · high confidence

New agent skills for data sourcing, factor research, and risk analysis

Added documentation and supporting scripts for several new agent skills: the \akshare\ and \ccxt\ skills provide free, keyless financial and cryptocurrency market data with built-in loader fallbacks; the \alpha-zoo\ skill enables browsing and benchmarking prebuilt factor libraries (e.g., Kakushadze 101, GTJA 191) with strict no-lookahead constraints; the \correlation-regime\ skill detects market fusion states and attributes crises using causal, hysteresis-based regime detection; and the \ashare-pre-st-filter\ skill predicts A-share delisting risks by combining financial data with regulatory penalty records fetched via a dedicated, SSRF-protected script. Also in this area: 4 commits (4 fixes) fixing agent; 21 commits (5 fixes) modifying agent; Fixes to market data loaders, optimizer logic, and agent replay visibility.

agent/src/skills · medium confidence · unverified

New agent tools for alpha research, backtest analysis, and market data

The agent now includes a suite of new tools in the \agent/src/tools\ directory to support quantitative research and market data access. The Alpha Zoo tools (\alpha\_zoo\, \alpha\_bench\, \alpha\_compare\) allow users to browse, benchmark, and compare alpha strategies across universes like CSI300 and S&P 500, with results paginated to fit context limits. The \backtest\_summary\ tool provides structured, JSON-safe summaries of backtest runs, including equity previews and metrics. New market data tools include \block\_trades\_tool\ for A-share block trades, \br\_market\_data\_tool\ for Brazilian market data, and \cashflow\_analytics\_tool\ for portfolio performance analysis. Safety improvements include \\_shell\_safety\ to prevent broad Python process termination and \\_result\_paging\ to ensure tool results are not silently truncated.

agent/src/tools · high confidence

New backtest artifacts, constraints, and correlation infrastructure

This change introduces several new components to the backtest engine. It adds Binance USD-M drift evidence artifacts (binance\_shadow\_evidence.py, perpetual\_evidence.py) for deterministic reconciliation and audit trails. It introduces a composable weight constraints layer (constraints.py) allowing per-name caps/floors and group exposure limits. It adds a cross-asset correlation matrix computation (correlation.py) and a correlation-regime timeline endpoint (regime.py) using edge density and hysteresis. It also includes a loader health lane (loader\_health.py) for public-source connectivity checks, rebalance notes (rebalance\_notes.py) for turnover/weight-drift reporting, and a rebalance mask (rebalance\_mask.py) for calendar-triggered scheduling.

agent/backtest · high confidence

New backtest infrastructure: benchmark resolution, USD-M reconciliation, factor costs, and options payoff math

This change introduces a suite of new modules that expand the backtesting engine's capabilities. The \benchmark.py\ module provides ticker resolution and fetch logic for market-specific benchmarks (e.g., SPY for US equities, CSI 300 for A-shares) to ensure accurate performance comparison. \binance\_account\_reconciliation.py\ and \binance\_tolerance\_calibration.py\ add offline, deterministic tools to compare local risk state against Binance USD-M exchange observations and calibrate tolerance bounds for those comparisons. \factor\_costs.py\ implements a trading-cost model for factor backtests, pricing capacity (ADV caps), market impact, and borrow costs in weight space. \options\_payoff.py\ adds deterministic math for multi-leg option strategy payoffs, including breakeven and max profit/loss calculations. Additionally, \tushare\_fundamentals.py\ and \rsshub\_events.py\ introduce new data loaders for point-in-time-safe fundamental data and event/sentiment feeds, while \engines/base.py\ and \models.py\ provide the shared execution loop and immutable data structures (Position, FillRecord, TradeRecord) that underpin these new capabilities. Also in this area: 7 commits (5 fixes) fixing vibe-trading-ai; 12 commits (2 fixes) modifying vibe-trading-ai; New research and analytics tools with hardened security and result paging.

vibe-trading-ai · medium confidence · unverified

New benchmarking and reporting scripts for alpha performance analysis

Added three new scripts in agent/scripts to support alpha zoo benchmarking and reporting: bench\_performance.py provides a development-only tool to benchmark factor operators and equity calculation performance; w4a\_run\_benches.py acts as a driver to run benchmarks across four zoo/universe combinations (gtja191/CSI300, alpha101/SP500, qlib158/CSI300, alpha101/BTC) and outputs summary JSON and HTML reports; w4a\_patch\_blog.py patches HTML blog posts with benchmark results, replacing TBD placeholders with actual statistics, theme survival data, and top/dead alpha cards. These scripts enable users to measure performance improvements and generate documentation from benchmark runs. Also in this area: 14 commits (10 fixes) fixing agent.

agent/scripts · medium confidence · unverified

New chat components for live trading, research goals, and agent activity

The chat interface now includes several new components to support advanced agent capabilities. ActivityLine displays real-time agent status, including live reasoning whispers and step counts. Composer handles multi-file uploads and integrates with live trading and goal panels. GoalPanel allows users to track and edit research goals with progress indicators. LiveRuntimePanel manages live trading status, including mandate proposals and broker connections. MandateProposalCard presents trading mandate options for user approval. ModelRuntimeBar shows current LLM provider and reasoning effort. PineScriptViewer displays TradingView Pine Script code. ProgressBar provides visual progress tracking. RunnerStatus shows broker connection and runner status. ScheduledResearchProposalCard handles scheduled research task approvals. SwarmStatusCard displays swarm agent status. Also in this area: 9 commits (2 fixes) modifying frontend; Fixes for scheduled delivery editing, email format selection, and report downloads.

frontend/src/components/chat · medium confidence · unverified

New contributor documentation and packaging infrastructure

The repository now includes an Agent Contributor Guide (\AGENT\_CONTRIBUTOR\_GUIDE.md\) detailing safety boundaries, high-risk surfaces, and targeted test commands for AI-assisted contributions, alongside a comprehensive \CONTRIBUTING.md\ covering the Developer Certificate of Origin (DCO) sign-off requirements, Alpha Zoo factor review checklists, and broker bring-up procedures. A \MANIFEST.in\ file ensures that documentation, tests, and environment examples are correctly bundled in source distributions, while \CODE\_OF\_CONDUCT.md\ and \NOTICE\ files establish community standards and third-party attribution.

(repo-wide) · high confidence

New dedicated pages for Alpha Zoo, Correlation, Options Lab, Portfolio, Reports, Runtime, Scheduled tasks, and Settings

The application now includes a suite of new dedicated frontend pages to expose advanced capabilities. The Alpha Zoo page allows users to browse, filter, and benchmark alpha factors across different universes. A new Correlation page provides a correlation matrix and regime timeline for asset analysis. The Options Lab enables users to build options strategies, view payoff diagrams, and analyze Greeks. The Portfolio page offers a multi-account overview with source management and history. The Reports page serves as a library for backtest results with filtering and sorting. The Runtime page provides a live status dashboard for brokers and mandates. The Scheduled page allows users to create and manage timezone-aware research schedules. Finally, the Settings page consolidates configuration for LLM providers, data sources, and API keys.

frontend/src/pages · high confidence

New front-end chart components for attribution, factor research, options, and correlation analysis

The charts directory now includes a suite of new React components that bring attribution, factor research, options, and correlation visualizations to the user interface. Users can now view Brinson attribution via cumulative and waterfall charts, explore factor research through IC series and group equity plots, and analyze options strategies using payoff diagrams and scenario heatmaps. Additionally, the interface supports correlation matrix heatmaps, regime timeline density plots, annual and monthly returns heatmaps, Monte Carlo path simulations, and statistical distribution charts, all styled with the application's theme and localized via i18n.

frontend/src/components/charts · high confidence

New frontend library modules for security, data processing, and UI logic

This change introduces a suite of new utility modules in the frontend library to support new product capabilities and improve robustness. Security is hardened with \apiAuth.ts\ (SSE ticket exchange to keep API keys out of URLs) and \escapeHtml.ts\ (XSS prevention for chart tooltips). Data handling is expanded with \positions.ts\ (parsing and classifying multi-market positions), \tearsheet.ts\ (computing monthly/annual returns and drawdowns), \cadence.ts\ (parsing scheduled-run cron expressions), and \attachments.ts\ (managing file attachments in prompts). UI and configuration logic is added via \options.ts\ (Options Lab types and validation), \swarmStatus.ts\ (normalizing swarm agent states), \theme-store.ts\ (centralized dark-mode state), and \tools.ts\ (tool name localization). The \api.ts\ module is significantly expanded with new interfaces for portfolio snapshots, correlation regimes, and report downloads, while \echarts.ts\ registers additional chart types (Heatmap, Pie, Treemap) and \formatters.ts\ adds Chinese (zh-CN) metric labels. Also in this area: 1 commit (1 fix) fixing frontend; 9 commits (2 fixes) modifying frontend; Fixes for scheduled delivery editing, email format selection, and report downloads; Users can now choose HTML or PDF formats for scheduled email reports; Edit scheduled runs and refine delivery target selection.

frontend/src/lib · medium confidence · unverified

New fundamental alpha factors for US, CN, and HK equities

The fundamental factor zoo now includes a comprehensive set of new alpha factors for US, CN, and HK equities, all designed to be point-in-time (PIT) safe. The update introduces quality factors (accruals, cash-to-assets, CFO yield, gross profitability, operating margin, ROA, ROE), value factors (book-to-market, earnings yield, FCF yield, dividend yield, P/B, P/S), and growth factors (asset growth, earnings growth, revenue growth). For the China (CN) universe, the release also adds industry-neutralized variants for dividend yield, P/B, and P/S to mitigate sector bias. These factors are computed daily and are available for use in the agent's strategy logic.

agent/src/factors/zoo/fundamental · high confidence

New fundamental data loader with SEC XBRL schema and A-share price conversion

This change introduces the foundational components for a new fundamental data loading system. It adds a unified schema (\_fundamental\_schema.py) that maps SEC XBRL concepts to financial fields (revenue, cogs, gross\_profit, etc.) and defines derived metrics like ROE and asset\_growth using a calendar-based lag. It also introduces a new HTTP helper module (\_http.py) for per-host request throttling and session reuse, a symbol utility (\_symbol\_utils.py) for detecting ETF prefixes, and a conversion module (additive\_conversion.py) to transform additive A-share price adjustments into multiplicative ratios. These components are integrated into the AKShare loader (akshare\_loader.py) to support A-share fundamental data and proper price adjustment handling.

agent/backtest/loaders · high confidence

New guided configuration panels for channels, data sources, and QVeris in Settings

The Settings page now includes dedicated, guided configuration interfaces for several key areas. Users can configure IM channels (DingTalk, QQ, Email, WebSocket, Feishu) with per-channel setup guides, connection testing, and hot-toggle enable/disable controls. A new Source Priority card allows reordering data-source fallbacks per market (including Argentina) with immediate backend application. Additionally, a QVeris settings panel enables integration with the paid data marketplace, supporting both free and paid modes, API key management, and budget credit configuration. Also in this area: 1 commit (1 fix) fixing frontend; 9 commits (2 fixes) modifying frontend; Fixes for scheduled delivery editing, email format selection, and report downloads.

frontend/src/components/settings · medium confidence · unverified

New institutional research slash commands

The CLI now exposes six new slash commands for institutional research workflows: /comps (comparable company analysis), /dcf (discounted cash flow valuation), /attrib (Brinson-Fachler performance attribution), /memo (investment memo), /earnings (earnings surprise bridge), and /screen (systematic idea screen). These commands provide structured, step-by-step execution skeletons with worked numeric examples and explicit missing-data policies, ensuring that analyses are grounded in retrieved data rather than model memory. Each command is implemented as a thin handler module that delegates to a shared runner, which constructs detailed prompts for the agent and handles argument parsing and length capping.

agent/cli/commands/institutional · high confidence

New interactive CLI UI with banner, rail dashboard, and transcript rendering

The CLI now includes a new visual interface module that renders a startup banner with a gradient logo, displays agent activity in a codex-style rail dashboard, and formats assistant responses as rich tables while stripping horizontal rules from markdown output.

agent/cli/ui · high confidence

New local development workflow script

A new \scripts/dev\ bash script has been added to streamline the local development environment. It provides commands to start, stop, restart, and check the status of the backend and frontend services, automatically managing process IDs and logs. The script handles environment setup, including locating the Python binary and installing frontend dependencies, and waits for services to be ready before completing.

scripts · high confidence

New modular trading connector framework with read-only broker support

The trading subsystem has been restructured into a modular connector architecture, introducing a profile-based registry and a capability matrix that is automatically generated into the README. This change adds read-only connectors for Scalable Capital (via Agentic MCP) and Trading 212 (via public REST API), alongside a new connection management system that supports account-scoped reads and credential storage via the OS keyring. The framework also introduces local plugin discovery for user-installed read-only connectors and machine-readable onboarding contracts to streamline the setup flow for various broker integrations.

agent/src/trading · high confidence

New multi-IM channel adapter layer with secure configuration and connection testing

The agent now includes a new plugin-based architecture for connecting to chat platforms (Telegram, Discord, Slack, Feishu, WhatsApp, DingTalk, QQ, and Email) via a dedicated \agent/src/channels\ module. This introduces a \ChannelManager\ and \MessageBus\ to decouple inbound messages from the agent loop and handle outbound routing. Key capabilities include per-channel connection tests (e.g., DingTalk, Discord, Email) that validate credentials without starting the adapter, and a robust configuration system with fail-safe secret masking to prevent credentials from being exposed in the Web UI. The system also supports hot-reloading of channel configurations without full restarts, streaming status updates to users, and secure handling of inbound media and outbound replies.

agent/src/channels · high confidence

New provider infrastructure and resilience features

This change introduces a new provider management layer and resilience helpers. It adds a \claude-cli\ provider with a strict \python-only\ bash policy to limit tool execution, and a \content\_filter\ module that tolerates LLM content-moderation blocks by skipping iterations and using a circuit breaker. It also introduces a \CustomProviderStore\ for managing custom LLM profiles and secrets, and a \session\_context\ module to bind per-conversation session IDs (e.g., for OpenCode). The provider registry is expanded with new providers like OpenAI Codex (OAuth), OpenCode, SiliconFlow, NVIDIA NIM, Novita AI, and others, while the LLM factory is updated to normalize OpenAI Responses stream events and disable HTTP proxy routing for specific providers.

agent/src/providers · high confidence

New slash commands and CLI command structure

The CLI now supports a new set of slash commands accessible via the interactive interface, including /goal for managing finance research goals, /playbook for scheduled research templates, /strategy-evidence for refreshing strategy evidence caches, and /update for self-upgrading the tool. These commands are organized under a new package structure in agent/cli/commands, which includes modules for chat interactions, session management, help, memory, and institutional research workflows. The slash router has been updated to include these new commands in the typeahead and help screens, providing users with a more structured and discoverable way to interact with the agent's capabilities.

agent/cli/commands · high confidence

Options Lab panel with payoff diagram, spot-IV scenarios, Greeks and live chain

The WebUI now includes an Options Lab section that lets users build and analyze options strategies. Users can construct multi-leg strategies using a StrategyBuilder with preset templates, view real-time market data in an OptionsChainTable (showing strikes, bid/ask, volume, open interest, and implied volatility), and see risk metrics displayed in GreeksCards (Delta, Gamma, Theta, Vega, Rho) with color-coded tones. The interface is fully internationalized via react-i18next, supporting localized labels and descriptions for all options-related UI elements.

frontend · high confidence

Persistent live-trading runtime with crash-safe state and preemptive halt sweep

The live trading runtime now persists its state across process restarts to prevent silent failures and duplicate actions. A crash-safe job store ensures the scheduler's job list survives SIGKILL or crashes without truncation, and quarantines corrupt files rather than starting with zero jobs. A new preemptive halt sweep module cancels all resting orders and, when permitted by the mandate, flattens open positions, with every broker call and error audited to a durable ledger. To prevent duplicate sweeps after a restart, the system persists a latch tied to the specific halt episode, using atomic file claims to ensure only one runner executes the sweep per episode, even in inter-process scenarios.

agent/src/live/runtime · high confidence

Turnover-aware optimizer and corrected backtest windowing

A new TurnoverAwareOptimizer is introduced, applying an L1 penalty on portfolio turnover to reduce rebalancing frequency, with support for per-asset and per-group exposure caps. The backtest engine now correctly restricts the optimization window to data strictly preceding the decision bar to prevent look-ahead bias, and passes returns to optimizers in position space (sign-adjusted for shorts) so that mean and covariance calculations reflect the actual long/short exposures being sized. Additionally, the RiskParity optimizer has been updated to enforce long-only constraints and uses a more robust optimization routine.

agent/backtest/optimizers · high confidence

Architecture

CLI refactored from a single file into a structured package

The agent CLI has been reorganized from a single legacy file into a modular package (agent/cli). The original implementation is preserved in cli/\_legacy.py to maintain backward compatibility, while the new package structure introduces dedicated modules for versioning (cli/\_version.py), input handling (cli/input.py), slash-command completion (cli/completer.py), and reusable UI components like chat logs and tool event renderers (cli/components). This change improves code maintainability and testability without altering the user-facing command-line interface.

agent/cli · high confidence

Behavioural changes

10 commits (5 fixes) modifying frontend/src/components/options

A change to existing behaviour in frontend/src/components/options — 10 commits (5 fixs), 9 files: “refactor: connect research data to calculations and carry evidence between roles”; “Improve UI responsiveness and task navigation”; “Fix analysis data fallback and recoverable research workflows”; and 7 more.

frontend/src/components/options · medium confidence · unverified

16 commits (6 fixes) modifying wiki/alpha-library

A change to existing behaviour in wiki/alpha-library — 16 commits (6 fixs), 6 files: “feat(wiki): Alpha Library auto-render + "Alpha 191 in 2026" blog”; “docs(data): clarify A-share source installation and fallback (\#2138)”; “fix(wiki): finish crawlability metadata, routes and PR checks”; and 13 more (refs \#2138, \#1802).

wiki/alpha-library · medium confidence · unverified

21 commits (5 fixes) modifying agent

A change to existing behaviour in agent — 21 commits (5 fixs), 189 files: “refactor: connect research data to calculations and carry evidence between roles”; “test: retain live report errors as offline evaluation cases”; “refactor(providers): unify in-process gateway and settings catalog”; and 18 more. Also in this area: Introduce offline harness for deterministic post-run evaluation.

(repo-wide) · medium confidence · unverified

23 commits (11 fixes) modifying agent/src/tools/paper\_trade

A change to existing behaviour in agent/src/tools/paper\_trade — 23 commits (11 fixs), 14 files: “fix(paper-trade): version volume-share matching as pt-rules-v5”; “feat(paper-trade): add the pt-v1.3 order contract”; “fix(paper-trade): preserve volume ratio precision”; and 20 more (refs \#1939, \#2176).

_agent/src/tools/paper\trade · medium confidence · unverified

9 commits (1 fix) modifying wiki/tutorials

A change to existing behaviour in wiki/tutorials — 9 commits (1 fix), 3 files: “docs(data): clarify A-share source installation and fallback (\#2138)”; “fix(wiki): finish crawlability metadata, routes and PR checks”; “docs: add Chinese beginner tutorial”; and 6 more (refs \#2138).

wiki/tutorials · medium confidence · unverified

9 commits (2 fixes) modifying wiki/scripts

A change to existing behaviour in wiki/scripts — 9 commits (2 fixs), 7 files: “feat(wiki): Alpha Library auto-render + "Alpha 191 in 2026" blog”; “docs(data): clarify A-share source installation and fallback (\#2138)”; “fix(wiki): finish crawlability metadata, routes and PR checks”; and 6 more (refs \#2138).

wiki/scripts · medium confidence · unverified

Context compaction now uses real token budgets instead of fixed estimates

The agent's context management has been refactored to size compaction based on the model's actual context window and real token usage rather than a fixed 40K estimated token threshold. This change prevents the loss of evidence in long-running sessions by dynamically adjusting memory pressure based on the specific model's capabilities (e.g., 272K for gpt-6-sol) and introduces a five-layer compaction system that includes zero-cost text folding and protected summary tails. Additionally, the agent now provides structured progress updates and heartbeats for long-running tools, tracks tool call failures to prevent infinite retry loops, and enforces stricter output principles requiring every reported number to be tied to a specific tool call.

agent/src/agent · high confidence

Enhanced theme persistence and robust SSE connection management

The frontend now persists the dark/light theme preference to local storage and synchronizes it across browser tabs and system preference changes, ensuring a consistent user experience. Additionally, the Server-Sent Events (SSE) hook has been refactored to handle authentication tickets for secure connections, ignore stale connections during retries, and support a wider range of event types including new swarm, goal, and live trading signals, improving reliability and feature coverage.

frontend/src/hooks · high confidence

Improved session reliability with streaming checkpoints and corruption resilience

Sessions now persist partial assistant responses during streaming via durable checkpoints, ensuring that interrupted or crashed runs recover their last known text rather than losing it entirely. The session store has been hardened to gracefully skip corrupt session.json files, malformed JSONL message lines, and invalid attempt records without aborting the entire listing or read operation. Additionally, the service enforces a strict one-run-per-session rule to prevent message interleaving and recovers interrupted attempts on restart by marking them with a clear user-facing status.

agent/src/session · high confidence

Improved state persistence reliability and simplified run lifecycle management

The run state store now ensures data integrity by using fsync when writing the state.json file, preventing corruption or truncation in the event of a crash. Additionally, the module has been streamlined to focus on directory creation and status recording; specific artifact persistence methods (such as saving planner outputs, design specs, and RAG results) have been removed, and a new method allows runs to be explicitly marked as cancelled by the user, distinguishing user-initiated stops from system failures.

agent/src/core · high confidence

Introduce security compatibility helpers for channel adapters

Added a new security module containing compatibility helpers that bridge existing channel logic with the agent's security layer. This includes re-exporting URL validation functions (validate\_url\_target, validate\_resolved\_url) and path containment checks (is\_path\_within) from the channels utilities, defining a WorkspaceScopeError for access violations, and providing a package init for tool output sanitization metadata.

agent/src/security · high confidence

Layout overhaul: expanded navigation, improved accessibility, and robust connection handling

The main layout component now features a significantly expanded sidebar with new navigation entries for Runtime, Scheduled, Reports, Portfolio, Alpha Zoo, Options Lab, Settings, and Correlation, all fully localized via react-i18next. Accessibility has been improved with ARIA labels, skip-to-content links, and proper icon hiding. The ConnectionBanner has been refactored to only display during active reconnection attempts (hiding the initial disconnected state), includes a manual reload button for persistent disconnections, and uses semantic ARIA roles. Sidebar collapse state is now persisted using a shared storage utility that supports cross-tab synchronization, and session lists automatically refresh on external title updates.

frontend/src/components/layout · high confidence

Refactor CLI structure and expand MCP tool surface

The CLI entry point has been restructured into a modular package (agent/cli) with new pure formatting and spinner helpers, while the MCP server now exposes 75 tools (up from 16) including research goals, strategy discovery, and extensive market data and fundamental analysis endpoints. The server also introduces a Streamable HTTP transport for modern clients and enforces explicit opt-in for shell tools to prevent remote code execution.

agent · high confidence

Visual identity refresh and dark mode support

The application's visual identity has been updated with a new logo and favicon featuring an orange-to-yellow gradient and a bar-chart icon, replacing the previous blue network-node design. The brand name has also changed to "Vibe-Trading". Additionally, a new theme-boot script has been added to automatically apply a dark or light theme based on user preference or system settings, ensuring consistent styling across the interface.

frontend/public · high confidence

Web UI overhaul with route-level lazy loading and new pages

The frontend application has been restructured to significantly improve initial load performance and expand feature coverage. Route-level lazy loading has been implemented for all major pages (Home, Agent, RunDetail, Compare, Settings, Runtime, Scheduled, Reports, Portfolio, Correlation, OptionsLab, AlphaZoo), reducing the initial bundle size and introducing a loading state during navigation. The router has been updated to include several new pages: Settings, Runtime, Scheduled, Reports, Portfolio, Correlation, OptionsLab, and AlphaZoo (with sub-routes for bench, compare, and individual alpha IDs). The Home page is now accessible at /about, and the default route (/) now loads the Agent page. Additionally, the UI has been enhanced with a new dark mode surface system for better contrast, RTL layout support, streaming animations, and improved accessibility features like reduced-motion support.

frontend/src · high confidence

Fixes

1 commit (1 fix) fixing tools/wiki

A fix in tools/wiki — 1 commit (1 fix), 4 files: “fix(wiki): finish crawlability metadata, routes and PR checks”.

tools/wiki · medium confidence · unverified

14 commits (8 fixes) fixing agent/src/research\_artifacts

A fix in agent/src/research\_artifacts — 14 commits (8 fixs), 7 files: “fix(research): isolate exports and bind asynchronous revision views”; “fix(research): isolate raw HTML in artifact exports”; “feat(research): export complete artifact series”; and 11 more (refs \#2164).

_agent/src/research\artifacts · medium confidence · unverified

22 commits (14 fixes) fixing agent/src/data\_gateway

A fix in agent/src/data\_gateway — 22 commits (14 fixs), 18 files: “fix(gateway): align skipped-route retries with source recovery deadline”; “fix(gateway): resume unfinished history on application restart”; “test: record complete mixed gateway scan and memory regressions”; and 19 more. Also in this area: 21 commits (5 fixes) modifying agent.

_agent/src/data\gateway · medium confidence · unverified

23 commits (18 fixes) fixing agent/src/factors/zoo/gtja191

A fix in agent/src/factors/zoo/gtja191 — 23 commits (18 fixs), 193 files: “fix(factors): gtja191\_098 keeps NaN through an unknown drift; the smoother exception to the NaN contract is stated and pinned (\#1463)”; “fix: complete September 27 reviews and surface verified run-card evidence”; “fix(factors): fix gtja191\_038 NaN propagation”; and 20 more (refs \#1463, \#17, \#101, \#55, \#170, …).

agent/src/factors/zoo/gtja191 · medium confidence · unverified

Agent market data routing and UI service fixes

The agent's market data handling now uses a dedicated symbol-to-source router (agent/src/market\_data.py) that explicitly routes UK equities (.L), Canadian venues (.TO/.V), Argentina (.BA), Korea (.KS/.KQ), and FX/suffix conventions to Yahoo, while keeping A-shares and HK equities on Tencent. This resolves issues where these symbols previously fell through to incompatible China-market loaders. The UI services (agent/src/ui\_services.py) have been updated to reconstruct charts via the loader registry, skip non-positive moving-average periods, cap CSV loads, and filter trade markers by symbol, ensuring the frontend receives consistent and valid data.

agent/src · high confidence

Swarm reliability, security, and grounding improvements

Swarm runs now include real-time market data grounding to prevent LLMs from citing stale training prices, with support for bare US tickers and configurable symbol caps. Error visibility is improved via a new serialization layer that consistently surfaces task errors across all read paths, and preset management now supports user-defined overrides from \~/.vibe-trading/swarm/presets/ alongside bundled defaults. Security and stability are enhanced with path-traversal validation for preset names and run IDs, redaction of internal paths in errors, Windows-safe atomic file writes with retry logic, and unique temporary file names to prevent race conditions during concurrent writes.

agent/src/swarm · high confidence

Test coverage

10 commits adding/updating tests in frontend/src/components/charts/\_\tests\\_; 3 commits adding/updating tests in agent/tests/harness\eval; 42 commits adding/updating tests in frontend/src/lib/\\tests\\; 5 commits adding/updating tests in agent/tests/factors/fixtures; 7 commits adding/updating tests in agent/tests/quantlib/valuation; 75 commits adding/updating tests in frontend/src/pages/\\tests\\_; 83 commits adding/updating tests in agent/tests/quantlib; Added MockAdvisory for testing advisory integrations; Added comprehensive test coverage for Tier 2 memory features; Added factor test suite with NaN propagation and golden-output regression tests; Added frontend test infrastructure and test helpers; Added frontend test suite for chat components; Added i18n locale parity and content integrity tests; Added memory retrieval benchmark suite with quality gates; Added test coverage for common UI components; Added test coverage for the new scheduling subsystem; Added test coverage for useDarkMode and useSSE hooks; Added tests for Channel, Model Picker, and Source Priority settings components; Added tests for Vite proxy configuration; Added tests for agent store and streaming DOM stability; Added tests for the PositionsTab component; Added tests for the positions sectors API endpoint; Added unit and integration tests for the OpenBB Workspace agent bridge; Test infrastructure expansion for MCP transports and agent goal context.

Dependencies

Major dependency upgrades and new channel SDK support

The application has upgraded its core Python dependencies, including LangChain to 1.3.9, LangGraph to 1.2.5, and FastMCP to 2.14.0, while adding support for document readers (PDF, DOCX, XLSX, PPTX) and new communication channels (Feishu, Telegram) via a dedicated requirements file. The frontend has been upgraded to React 19.2, React Router 8, and Vite 8.2, and the Electron desktop build now targets Node 22 and Electron 43.1.1, with updated lock files to address security advisories.

(dependencies) · high confidence

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

How this codebase got here

Score

  • CAI 52 → 57 (+5.2)
  • Rubric changed (rubric-2026.11.2 → rubric-2026.11.4) — scores are not directly comparable.

Lenses

  • Code Health 66 → 64 (-1.5)
  • Architecture 92 → 93 (+1.4)
  • Maturity 60 → 63 (+3.4)
  • Readiness 42 → 49 (+7.9)
  • Security 78 → 78 (-0.1)
  • Accessibility 56 → 63 (+7.0)
  • Performance 60 → 60 (+0.0)

Resolved (203)

  • Change coupling: api.ts ↔ router.tsx (frontend/src/lib/api.ts)
  • ChannelRuntime._wait_for_reply (cognitive 17) (agent/src/channels/runtime.py)
  • Correlation.Correlation (cognitive 24) (frontend/src/pages/Correlation.tsx)
  • Correlation.Correlation (cyclomatic 25) (frontend/src/pages/Correlation.tsx)
  • DataLoader._fetch_one (cognitive 17) (agent/backtest/loaders/qveris_loader.py)
  • Duplicated block (10 lines × 2) (agent/src/agent/grounding/identity.py)
  • Duplicated block (10 lines × 6) (agent/src/trading/service.py)
  • Duplicated block (10–13 lines × 12) (agent/backtest/loaders/akshare_loader.py)
  • Duplicated block (10–13 lines × 2) (agent/src/channels/wecom.py)
  • Duplicated block (11 lines × 2) (agent/src/agent/grounding/figures.py)
  • Duplicated block (11 lines × 2) (agent/src/skills/tushare/scripts/fund_data_example.py)
  • Duplicated block (11 lines × 2) (agent/src/tools/get_fundamentals_tool.py)
  • Duplicated block (11 lines × 2) (agent/src/trading/service.py)
  • Duplicated block (12–14 lines × 4) (agent/backtest/loaders/alphavantage_loader.py)
  • Duplicated block (13–17 lines × 2) (agent/src/factors/cli_handlers.py)
  • Duplicated block (14 lines × 2) (agent/src/strategy_store/sqlite_store.py)
  • Duplicated block (14 lines × 4) (agent/backtest/loaders/fmp_loader.py)
  • Duplicated block (16 lines × 3) (agent/backtest/loaders/pykrx_loader.py)
  • Duplicated block (18 lines × 2) (agent/src/skills/okx-market/scripts/candle_data_example.py)
  • Duplicated block (19–20 lines × 2) (agent/src/providers/llm.py)
  • …and 183 more

New (524)

  • Correlation.CorrelationAnalysis (cognitive 24) (frontend/src/pages/Correlation.tsx)
  • Correlation.CorrelationAnalysis (cyclomatic 25) (frontend/src/pages/Correlation.tsx)
  • DataLoader._fetch_candles (cognitive 18) (agent/backtest/loaders/okx.py)
  • DataLoader.fetch (cognitive 18) (agent/backtest/loaders/akshare_loader.py)
  • DataLoader.fetch (cognitive 18) (agent/backtest/loaders/eastmoney_loader.py)
  • Duplicated block (10 lines × 2) (agent/src/agent/instrument_identity.py)
  • Duplicated block (10 lines × 5) (agent/backtest/loaders/fmp_loader.py)
  • Duplicated block (10–11 lines × 8) (agent/backtest/loaders/alphavantage_loader.py)
  • Duplicated block (11 lines × 2) (tools/gateway_full_scan_probe.py)
  • Duplicated block (11 lines × 3) (agent/backtest/loaders/akshare_loader.py)
  • Duplicated block (11–13 lines × 3) (agent/backtest/loaders/alphavantage_loader.py)
  • Duplicated block (12 lines × 2) (tools/gateway_full_scan_probe.py)
  • Duplicated block (12 lines × 2) (tools/gateway_minute_restart_probe.py)
  • Duplicated block (12 lines × 6) (agent/backtest/loaders/alphavantage_loader.py)
  • Duplicated block (12–13 lines × 2) (agent/src/tools/gildata_fin_query_tool.py)
  • Duplicated block (12–13 lines × 3) (agent/backtest/loaders/akshare_loader.py)
  • Duplicated block (13 lines × 3) (agent/backtest/loaders/alphavantage_loader.py)
  • Duplicated block (13 lines × 4) (agent/backtest/loaders/alphavantage_loader.py)
  • Duplicated block (16 lines × 4) (agent/backtest/loaders/fmp_loader.py)
  • Duplicated block (18 lines × 2) (agent/src/skills/okx-market/scripts/candle_data_example.py)
  • …and 504 more

Changes since last survey

  • 238 commits — 126 feature/other, 112 fixes

By area

  • agent/src — 107 commits
  • (repo) — 29 commits
  • agent/tests — 29 commits
  • (root) — 26 commits
  • frontend/src — 24 commits
  • agent/backtest — 5 commits
  • desktop/electron — 4 commits
  • guides/capabilities.md — 3 commits
  • assets/features — 2 commits
  • tools/gateway_soak.py — 2 commits
  • .github/workflows — 1 commit
  • agent/SKILL.md — 1 commit
  • agent/cli — 1 commit
  • agent/mcp_server.py — 1 commit
  • frontend/package-lock.json — 1 commit
  • tools/gateway_combined_report.py — 1 commit
  • tools/wiki — 1 commit

Notable commits

  • fix: Fix options refresh lifecycle, preset strikes and leg limits
  • fix: Merge November 12 external-review fixes and README news
  • fix: Merge PR #2186 with volatility surface coverage fix
  • fix: Merge pull request #2185 from Jackzigen/fix/multi-line-percent-headers
  • fix: Merge pull request #2192 from Jackzigen/fix/2126-bare-large-integers
  • fix: Merge pull request #2194 from divya0795/fix/channels-stale-reply
  • fix: Merge pull request #2195 from divya0795/fix/strict-bench-panel-enrichment
  • fix: fix(agent): keep retry directives compatible with Anthropic
  • fix: fix(alpha-bench): skip single-observation IC series in the strict bench
  • fix: fix(api): bound research-artifact evidence and id fields
  • fix: fix(attribution): degrade asset-class Brinson when a class member is unpriced
  • fix: fix(background): restore pending notifications without trusting corrupt records
  • fix: fix(backtest): accept Yahoo's .SS Shanghai spelling in the code check
  • fix: fix(backtest): buy the benchmark at the first evaluated open, as the strategy is
  • fix: fix(backtest): keep fast-path margin positive at a negative entry price
  • fix: fix(backtest): recognise bare codes for every China futures product the engine prices (#2177)
  • fix: fix(backtest): recognise bare three-digit Zhengzhou futures codes
  • fix: fix(backtest): refuse a code no market rule recognises
  • fix: fix(backtest): select and report the benchmark actually evaluated
  • fix: fix(backtest): show why the portfolio risk x-ray is unavailable
  • …and 218 more

Architecture

  • Unchanged — 0 containers · 1 contexts · 0 edges

Written by cca.canine.dev from the codebase's own history, inside the signed delivery this page is composed from.

Survey your own repository

HKUDS/Vibe-Trading was measured the same way every project in this corpus was: the same rubric, at a pinned commit, with the result published in full. Point a surveyor at a repository you know and see whether you agree with it.

About this page

  • The score is its most recent published measurement, taken on 16 November 2026 at a pinned commit. It is not a live figure and does not change until the project is measured again.
  • Measured at commit 1dbe5d515519565a47f0127b4e6d6fc4fbb5e4d2 — the exact code this score is about.
  • Scored under rubric-2026.11.4 — the same rubric and the same method as every other entry in this index.
  • Measured by Canine Code Analyzer (cca.canine.dev) using cca measure-2026.11.73+preprod-52de90b74216.