application.properties

spring.main.allow-bean-definition-overriding=true

# =============================================================================
# Browser4 Application Configuration
# =============================================================================
# This file documents ALL environment variables recognized by Browser4.
# Values below are PLACEHOLDERS and commented out by default.
# The CLI and backend read these from the OS environment, not from this file —
# this file serves as documentation and a local reference.
#
# To manage env vars interactively, use the helper scripts:
#   bash:   ./bin/env-manage.sh show|get|set|unset|export|list
#   pwsh:   .\bin\env-manage.ps1 show|get|set|unset|export|list
# =============================================================================

# ---------------------------------------------------------------------------
# CLI: Runtime & Paths
# ---------------------------------------------------------------------------

# Chrome extension auth token for auto-approval of attach requests.
# Find yours at: chrome-extension://.../connect.html → "Authentication Token" section
# The CLI reads this from the OS environment, not from this file.
# BROWSER4_EXTENSION_TOKEN=<your-real-token>

# Default session ID for the CLI (pre-selects a session so you don't need --session on every command)
# BROWSER4_CLI_SESSION=default

# Override CLI session state directory (default: ~/.browser4 on Linux/macOS, %APPDATA%/browser4 on Windows)
# BROWSER4_CLI_STATE_DIR=/custom/path/to/state

# Override runtime bundle & download cache directory
# BROWSER4_RUNTIME_DIR=/custom/path/to/runtime

# Override skills directory path (default: versioned install dir under RUNTIME_DIR)
# BROWSER4_SKILLS_DIR=/custom/path/to/skills

# Custom Chromium/Chrome browser executable path (auto-detected if unset)
# BROWSER4_BROWSER_PATH=/usr/bin/google-chrome

# JVM options injected into backend server launch, space-separated
# BROWSER4_SERVER_OPTS=-Dapp.name=browser4 -Xmx2g

# MCP-over-HTTP endpoint (opt-in; OFF by default). Enabling it binds a second
# endpoint (POST /mcp, default port 8088) and eagerly acquires a browser session
# at backend start, so it is only worth turning on for an MCP client such as
# Claude Desktop or Cursor. Pass it through the JVM options above, e.g.:
# BROWSER4_SERVER_OPTS=-Dmcp.http.enabled=true -Dmcp.http.port=8088
# mcp.http.enabled=true
# mcp.http.port=8088

# Directory for server startup logs (default: RUNTIME_DIR/logs)
# BROWSER4_SERVER_LOG_DIR=/path/to/logs

# ---------------------------------------------------------------------------
# CLI: Download & Mirror
# ---------------------------------------------------------------------------

# Override releases download base URL (default: GitHub Releases)
# BROWSER4_RELEASES_BASE_URL=https://mirrors.example.com/browser4/releases

# Override mirror configuration file path (default: runtime/mirrors.json)
# BROWSER4_MIRRORS_CONFIG=/path/to/custom-mirrors.json

# Mirror reachability check timeout in seconds (default: 5)
# BROWSER4_CLI_MIRROR_CHECK_TIMEOUT_SECS=10

# Per-mirror speed test download timeout in seconds (default: 30)
# BROWSER4_CLI_MIRROR_SPEED_TEST_TIMEOUT_SECS=15

# Cached mirror preference TTL in seconds (default: 86400 = 24 hours)
# BROWSER4_CLI_MIRROR_PREFERENCE_TTL_SECS=3600

# Download timeout in seconds (default: 1800 = 30 minutes)
# BROWSER4_CLI_DOWNLOAD_TIMEOUT_SECS=600

# Set to 1 to disable speed testing (falls back to TCP-only reachability)
# BROWSER4_CLI_DISABLE_MIRROR_SPEED_TEST=1

# Root search start directory for resolving relative paths (default: current working directory)
# BROWSER4_CLI_INVOKE_DIR=/home/user/projects

# ---------------------------------------------------------------------------
# CLI: Build & Bundle
# ---------------------------------------------------------------------------

# Set to 1/true/yes/on to force download from remote releases instead of local build
# BROWSER4_CLI_FORCE_REMOTE_BUNDLE=1

# Set to 1/true/yes/on to force rebuild runtime bundle from source
# BROWSER4_CLI_FORCE_REBUILD_BUNDLE=1

# ---------------------------------------------------------------------------
# CLI: HTTP Request Timeouts (seconds)
# ---------------------------------------------------------------------------

# Default HTTP request timeout (default: 30)
# BROWSER4_CLI_HTTP_TIMEOUT_SECS=60

# Navigation (page load) request timeout (default: 120)
# BROWSER4_CLI_NAVIGATION_TIMEOUT_SECS=180

# Text input request timeout (default: 90)
# BROWSER4_CLI_INPUT_TIMEOUT_SECS=120

# Snapshot request timeout (default: 60)
# BROWSER4_CLI_SNAPSHOT_TIMEOUT_SECS=90

# Agent tool request timeout (default: 180)
# BROWSER4_CLI_AGENT_TIMEOUT_SECS=300

# Act tool request timeout (default: 60)
# BROWSER4_CLI_ACT_TIMEOUT_SECS=120

# Crawl request timeout (default: 600)
# BROWSER4_CLI_CRAWL_TIMEOUT_SECS=1200

# ---------------------------------------------------------------------------
# CLI: Proxy
# ---------------------------------------------------------------------------

# Explicit CLI download proxy URL (also set by --proxy <url> on the CLI)
# BROWSER4_CLI_PROXY=http://proxy:8080

# Internal test flag (ignore)
# BROWSER4_CLI_TEST_TEMPORARY_PROFILE=0

# ---------------------------------------------------------------------------
# LLM / AI Provider API Keys
# ---------------------------------------------------------------------------
# Set ONE of these depending on which LLM provider you use.
# Priority order (in AgentToolExecutor): LLM_API_KEY → DEEPSEEK_API_KEY →
#   OPENROUTER_API_KEY → VOLCENGINE_API_KEY → OPENAI_API_KEY
#
# IMPORTANT: env-style keys (DEEPSEEK_API_KEY=...) below are only honored as
# REAL environment variables or -D system properties.  Inside a `.properties`
# file (including ~/.browser4/config/conf-enabled/*.properties) they do NOT
# bind to the engine's dotted lookups (deepseek.api.key) and extract/summarize
# would fail with "LLM API key is not configured".  At startup the backend
# automatically rewrites env-style LLM keys found in conf-enabled properties
# files to their dotted form (DEEPSEEK_API_KEY=... → deepseek.api.key=...),
# but the recommended file format is the dotted keys in the provider sections
# below.

# Generic LLM API key (fallback for all providers)
# LLM_API_KEY=sk-...

# OpenAI (env-var style; only valid as an environment variable)
# OPENAI_API_KEY=sk-...

# OpenRouter (env-var style; only valid as an environment variable)
# OPENROUTER_API_KEY=sk-or-v1-...

# DeepSeek (env-var style; only valid as an environment variable)
# DEEPSEEK_API_KEY=sk-...

# Volcengine (env-var style; only valid as an environment variable)
# VOLCENGINE_API_KEY=...

# Provider-specific Spring properties (uncomment the provider you want to use):

## OpenRouter
#openrouter.api.key=...
#openrouter.model.name=openai/gpt-5.4
#openrouter.base.url=https://openrouter.ai/api/v1/

#openrouter.api.key=...
#openrouter.model.name=bytedance-seed/seed-2.0-lite
#openrouter.base.url=https://openrouter.ai/api/v1/

## Volcengine (ByteDance)
# volcengine.api.key=...
# volcengine.model.name=doubao-seed-2-0-pro-260215
# volcengine.base.url=https://ark.cn-beijing.volces.com/api/v3

## DeepSeek
# deepseek.api.key=sk-...
# deepseek.model.name=deepseek-v4-flash
# deepseek.base.url=https://api.deepseek.com

# Per-request input length cap for the LLM chat model. PulsarRPA's
# ChatModelSettings defaults to 64_000 (a conservative floor across all
# providers) and silently truncates messages above it; deepseek-v4-flash
# supports a 1M-token context window, so raise the cap to match and leave
# headroom for output tokens.
llm.max.input.token.length=900000

## OpenAI-compatible
# openai.api.key=sk-...
# openai.model.name=gpt-4o
# openai.base.url=https://api.openai.com/v1

## Aliyun Qwen (DashScope, OpenAI-compatible)
# openai.api.key=sk-...
# openai.model.name=qwen-plus
# openai.base.url=https://dashscope.aliyuncs.com/compatible-mode/v1

# ---------------------------------------------------------------------------
# CAPTCHA Solving
# ---------------------------------------------------------------------------

# Enable automatic CAPTCHA detection and solving during browsing
# captcha.auto.solve.enabled=true

# Primary solving service: CAPSOLVER, TWO_CAPTCHA, ANTI_CAPTCHA
# captcha.service.provider=CAPSOLVER

# API keys — set as environment variables (preferred) or in application-private.properties
# captcha.capsolver.api.key=${CAPSOLVER_KEY}
# captcha.twocaptcha.api.key=${TWOCAPTCHA_KEY}
# captcha.anticaptcha.api.key=${ANTICAPTCHA_KEY}

# CAPSOLVER_KEY=<your-capsolver-key>
# TWOCAPTCHA_KEY=<your-2captcha-key>
# ANTICAPTCHA_KEY=<your-anticaptcha-key>

# Maximum time to wait for a solution (seconds)
# captcha.solve.timeout.seconds=120

# Interval between polling for task status (ms)
# captcha.poll.interval.ms=1000

# Enable automatic CAPTCHA detection on page load
# captcha.detection.enabled=true

# Types of CAPTCHA to auto-solve
# captcha.auto.solve.types=RECAPTCHA_V2,HCAPTCHA,TURNSTILE

# Report failed solves for refund (2Captcha / Anti-Captcha only)
# captcha.report.failed.enabled=true

# Maximum retry attempts for solving
# captcha.solve.max.retries=3

# ---------------------------------------------------------------------------
# Observability: OpenTelemetry + Metrics
# ---------------------------------------------------------------------------

# Enable/disable OpenTelemetry tracing (default: true)
# OTEL_TRACES_ENABLED=true

# OTLP gRPC collector endpoint (default: http://localhost:4317)
# OTEL_EXPORTER_OTLP_ENDPOINT=http://localhost:4317

# OpenTelemetry service name (default: browser4-agentic)
# OTEL_SERVICE_NAME=browser4-agentic

# OpenTelemetry service version (default: from build)
# OTEL_SERVICE_VERSION=4.8.1-SNAPSHOT

# Enable/disable metrics collection (default: true)
# METRICS_ENABLED=true

# Metrics name prefix (default: pulsar_agentic)
# METRICS_PREFIX=pulsar_agentic

# Comma-separated key=value pairs for metrics common tags
# METRICS_COMMON_TAGS=env=production,region=us-east-1

# ---------------------------------------------------------------------------
# Backend
# ---------------------------------------------------------------------------

# Comma-separated list of read actions whitelisted for tab operations
# (e.g. read_snapshot,read_state,read_cookies)
# BROWSER4_TAB_READ_ACTIONS_WHITELIST=read_snapshot,read_state

# Custom path to Playwright browser binaries
# PLAYWRIGHT_BROWSERS_PATH=/path/to/playwright-browsers

# ---------------------------------------------------------------------------
# Proxy (network)
# ---------------------------------------------------------------------------

# Proxy rotation URL for the browser
# proxy.rotation.url=https://your-proxy-provider.com/rotation-endpoint

# HTTP/HTTPS proxy environment variables (read by both CLI and backend):
#   http_proxy  / HTTP_PROXY   — HTTP proxy
#   https_proxy / HTTPS_PROXY  — HTTPS proxy
#   all_proxy   / ALL_PROXY    — Catch-all proxy

# ---------------------------------------------------------------------------
# E2E Testing
# ---------------------------------------------------------------------------

# External Browser4 service URL for E2E tests (skips local bundle build/launch)
# BROWSER4_E2E_SERVICE_URL=http://localhost:18182

# Alias for BROWSER4_E2E_SERVICE_URL
# BROWSER4_E2E_SERVER_URL=http://localhost:18182

# Host the Browser4 container uses to reach the fixture server (default: 127.0.0.1)
# BROWSER4_E2E_FIXTURE_HOST=host.docker.internal

# Override per-command timeout in E2E tests (seconds)
# BROWSER4_E2E_CLI_TIMEOUT_SECS=120

# Set to 1/true/yes/on for Maven-based startup in E2E tests
# BROWSER4_E2E_USE_MAVEN_STARTUP=1

# Force remote bundle download in E2E tests
# BROWSER4_E2E_FORCE_REMOTE_BUNDLE=1

# ---------------------------------------------------------------------------
# Mock Site Testing
# ---------------------------------------------------------------------------

# Port for mock site test server (default: 8182)
# MOCK_SITE_PORT=9090

# Wait seconds for mock site startup
# MOCK_SITE_WAIT_SEC=30

# ---------------------------------------------------------------------------
# Example Crawler Credentials (development only)
# ---------------------------------------------------------------------------

# Taobao login credentials for example scrapers
# PULSAR_TAOBAO_USERNAME=your-username
# PULSAR_TAOBAO_PASSWORD=your-password

# Simuwang login credentials for example scrapers
# PULSAR_SIMUWANG_USERNAME=your-username
# PULSAR_SIMUWANG_PASSWORD=your-password

# ── Agent feedback-loop overhaul (2026-08-21) ──────────────────────────────
# Native tool calling is the DEFAULT now; tool results feed back into the
# conversation automatically. Explicit fallbacks: `text`, `chat`.
# agent.tool.expose.mode=toolCalling
# Tiered task-adaptive tool disclosure: `tiered` (default) or `full` (flat list).
# browser4.agent.toolDisclosure=tiered
# Abort after N consecutive text-only responses (0 = disabled).
# browser4.agent.textOnlyStallLimit=5
# Finish-report gate validation: `strict` (default) or `warn`.
# browser4.agent.finishGateCheck=strict
# Max model<->tool round-trips inside one native tool-calling turn.
# browser4.agent.toolLoop.maxIterations=12

# Agent execution engine: `cli` (default, CLI tool-loop driving browser4-cli).
# `observe-act` (legacy ARIA-snapshot engine) is DEPRECATED — kept only for
# backward compatibility; new runs should stay on the default `cli`.
browser4.agent.runEngine=cli