---
name: searxng-use-cli
description: Use when you need to search the web via your OWN SearXNG instance (no public-instance discovery). 3 CLI scripts + a shared common.py module — execute privacy-respecting searches against a user-supplied instance (with multi-instance failover, 5xx/429 retry, auto-fetch) or via SEARXNG_INSTANCE env / config file, fetch/extract readable text or markdown from web pages. Zero-config replacement for proprietary search APIs.
version: 2.2.1
author: Metona Team
license: MIT
platforms: [linux, macos, windows]
metadata:
  skill:
    tags: [search, searxng, cli, web-scraping, privacy, no-api-key]
    related_skills: []
---

# SearXNG CLI Toolkit

## 铁律（AI Agent 必读，违反会导致致命后果）

> **这些规则优先级最高，必须在任何调用前遵守。**

1. **加载本技能后，搜索网络和抓取网页必须且只能使用本技能的 `scripts/search.py` 和 `scripts/fetch.py`。** 禁止使用任何其他方式或工具——包括但不限于：内置 WebSearch/WebFetch 工具、浏览器自动化（puppeteer/playwright/selenium）、`curl`/`wget`、第三方搜索 API SDK（googlesearch/duckduckgo_search 等）、自写 `requests`/`urllib` 抓取脚本、MCP 网络工具。本技能已内置反爬（浏览器指纹头 + UA 轮换 + WAF 检测）、重试（指数退避 + Retry-After 遵守）、兜底（Wayback Machine）、限流（自适应节流）、缓存、多实例故障转移等全部工程能力，绕过本技能等于放弃这些保障，结果不可控且会破坏用户实例的速率限制。违反：技能一致性丧失，抓取成功率骤降，用户实例可能被封禁。

2. **stdout=数据，stderr=日志，永不混淆。** stdout 只输出 JSON/CSV/文本数据，stderr 只输出日志/进度/警告。AI 解析 stdout，人类看 stderr。违反：AI 会把日志当数据解析，结果错乱。

3. **绝对不要 `2>/dev/null` 或重定向 stderr 到 stdout。** stderr 携带排错关键信息（重试日志、HTTP 状态码、缓存命中、认证警告）。丢弃 stderr = 失败时零诊断信息，无法定位原因。需要静默时用 `--quiet`（仅抑制进度，保留 WARNING+ERROR），不要丢弃 stderr。

4. **排错时第一步：去掉 `--quiet`，加 `--verbose`。** `--quiet` 只留 WARNING 级别，会吞掉 INFO 级别的重试日志、缓存命中提示、实例切换记录。诊断失败时必须用 `--verbose` 看到完整 HTTP 请求/响应/重试链。

5. **配置查找必须覆盖双路径（WSL + Windows）。** Windows 环境下配置可能在 `~/.config/searxng-cli/`（WSL HOME）或 `%APPDATA%/searxng-cli/`（Windows APPDATA）。检查配置存在性时两个路径都要查，否则会误判"无配置"并反问用户。

6. **实例 URL 必填，公共实例发现已移除。** 必须通过 `-i`、`SEARXNG_INSTANCE` 环境变量、或配置文件提供实例 URL。无实例时 `search.py` 报 `E_CONFIG` 退出，不要尝试猜测或硬编码公共实例。

## Overview

SearXNG is a privacy-respecting metasearch engine that aggregates results from 70+ search services without tracking users. This skill provides standalone Python CLI scripts — works with **any AI agent** or directly from your terminal.

**Public-instance discovery has been removed.** You must supply your own SearXNG instance URL (self-hosted or one you trust). This makes behavior deterministic and avoids depending on volatile public instances.

**Key capabilities:**

**Search & results**
- Multi-instance failover with parallel probing (faster failover, deterministic output order)
- Exponential-backoff retry on transient errors (403/429/5xx/connection) via shared retry-policy components in `common.py` (backoff/retryable-status/Retry-After), with the retry loop in each script
- `--verify` health-check mode (reachability / JSON-API / latency / POST / engine list / auth status)
- Cross-engine result deduplication (default on; `--no-dedup` disables) — collapses duplicate URLs ignoring tracking params (`utm_*`, `gclid`, etc.) and fragments
- Similarity deduplication (`--similarity-dedup`; off by default) — title-based SimHash + Jaccard, auto-skipped when results > 500
- Result sorting (`--sort-by {score,date,engine,none}`; default: score descending) — applied after dedup, before `--max-results`
- Domain allowlist/blocklist (`--include-domain` / `--exclude-domain`) — case-insensitive, ignores leading `www.`, exclude wins on conflict
- Pagination aggregation (`--pages N`) — fetch N pages in one run, merge with cross-page dedup
- Batch mode (`--queries-file`) — run multiple queries from a file in sequence, combined output

**Output formats**
- JSON (default, rich metadata), brief (title+URL+snippet), urls (plain list), CSV (spreadsheet-ready)
- Enhanced Markdown conversion — nested ordered lists, mixed `ul`/`ol` nesting, `<dl>`/`<dt>`/`<dd>` definition lists, GFM tables, fenced code blocks, blockquotes
- Structured JSON error output (in `--format json` mode) with `error_code` field for machine-readable failure reporting
- JSON Lines streaming (`--stream`) — each result emitted as a separate JSON line to stdout, enabling incremental processing
- Progress events (`--progress`) — structured JSON Lines events to stderr for real-time execution tracking, includes `request_id` for correlation

**Caching & config**
- SQLite result caching (`--cache-ttl`) with size cap + LRU eviction (`--cache-max-size`, default 100MB) — identical queries within a TTL skip the network entirely; `--clear-cache` / `--cache-stats` manage it
- Config file (`searxng.toml`) pre-sets most flags; `--config FILE` loads a non-default config; `instances.txt` for plain URL lists; `--save-config FILE` writes current args to a config file
- Instance resolution priority: `-i` → `SEARXNG_INSTANCE` env → config file (`./searxng.toml` → `~/.config/searxng-cli/searxng.toml` → `%APPDATA%/searxng-cli/searxng.toml` on Windows) → `instances.txt` (`./instances.txt` → `~/.config/searxng-cli/instances.txt` → `%APPDATA%/searxng-cli/instances.txt`)

**Network & auth**
- Proxy support (`--proxy`) for both search and fetch (sets `HTTP_PROXY`/`HTTPS_PROXY`/`NO_PROXY`)
- Auth via CLI flag, file, config file, or env var (`--auth-bearer` / `--auth-basic` + `*-file` variants) to avoid leaking secrets in shell history
- Config-file auth — `auth_basic` and `auth_bearer` fields in `searxng.toml` let AI agents set credentials once (priority: CLI > file > config > env)
- Credentials-file permission warning — `--auth-*-file` warns on stderr if the file is group/other-readable (POSIX only)

**Anti-bot & fetch stability**
- **Browser fingerprint headers** — `build_browser_headers()` sends full Sec-Ch-Ua / Sec-Fetch-* / Accept-Language / Accept-Encoding, not just User-Agent. Bypasses 80%+ of lightweight WAFs (Sec-Ch-Ua/Sec-Fetch-* only sent for Chrome/Edge; Firefox omits them). **Smart Accept-Encoding**: only advertises `br` when the brotli decompressor is actually installed, preventing servers from returning Brotli-compressed bytes that requests can't auto-decompress (which would produce garbled output)
- **15-UA pool** — Chrome 138-140 (Win/mac/Linux), Edge 138 (Win/mac), Firefox 140 (Win/mac/Linux), Safari 18 (mac). `get_ua_for_domain()` deterministically assigns one UA per domain (stable within a session, reproducible across processes)
- **requests.Session reuse** — module-level Session with connection pool + cookie persistence + TLS session resumption
- **Split timeouts** — `timeout=(connect, read)` tuple (connect capped at 10s, read defaults to 15s)
- **Retry-After compliance** — 429/503 responses read the `Retry-After` header (numeric seconds or HTTP date) and wait at least that long before retrying
- **Capped backoff** — `compute_backoff_delay()` caps at 60s
- **WAF fingerprint library** — `_detect_anti_bot()` (in `search.py`, used by `--fetch` flow) identifies Cloudflare / Imperva / PerimeterX / DataDome / Akamai / Anubis / generic challenges via `<title>` tag matching + full-document scan. Returns `waf_type` for AI-agent decisioning
- **Wayback Machine fallback** — 404/403/timeout AND anti-bot-blocked pages automatically retry via `https://web.archive.org/web/2/<url>` (latest snapshot). Default ON; `--no-fallback` disables. Wayback timeout = `min(--timeout, 10s)`; Wayback retries = `min(--retries, 2)`
- **Hard-blocked domain fallback** — `is_hard_blocked_domain()` (in `common.py`) detects known strong-anti-bot sites (www.baidu.com, baike.baidu.com, zhihu.com, weibo.com, mp.weixin.qq.com, douban.com, etc.) that almost always return 403. These trigger Wayback immediately on failure. List maintained in `HARD_BLOCKED_DOMAINS`. Note: baidu uses exact subdomain matching (www/baike/zhidao/tieba/wenku) to avoid over-blocking pan.baidu.com; zhihu/weibo/douban use subdomain wildcard matching
- **Adaptive throttling** — `AdaptiveThrottle` (in `search.py`) state machine: consecutive failures → double delay + halve concurrency; consecutive successes → gradual recovery; 429 → global pause. Thread-safe. Parameters configurable via `--throttle-failure-threshold` / `--throttle-pause-seconds` / `--throttle-max-delay`
- **readability-lite extraction** — when `<article>`/`<main>`/`role="main"`/content-class `<div>` are all missing and only `<body>` remains, picks the highest text-density `<div>`/`<section>`/`<article>` node (scored by text density + `<p>` count weighting). Text-density threshold is language-aware — CJK content uses 100 chars, other languages use 200 chars
- **PDF/document parsing** — `fetch.py` parses PDF (via `pdftotext` subprocess) and `.docx`/`.xlsx` (via stdlib `zipfile`). Unsupported binary types return `E_UNSUPPORTED_MEDIA`
- **`--fetch-report [json]`** — structured per-URL report to stderr after `--fetch N`: status, WAF type, fallback used, char count, plus throttle stats. `--fetch-report json` outputs a full JSON report (items array + summary)
- **`--referer`** — set Referer header for fetch requests (defaults to the instance URL when fetching result pages)
- **`--request-delay`** — configurable delay between fetch requests (default 0.3s; adaptive throttling may increase this on failures)
- **`search.py --fetch` output fields** (in `fetched` array): `anti_bot_detected` (bool), `waf_type` (str|null), `fallback_used` (str|null), `error_code` (str|null — structured error code on fetch failure)
- **`fetch.py` FetchResult fields** (standalone script output): `content` (str), `content_type` (str), `final_url` (str), `truncated` (bool), `user_agent` (str), `error_code` (str|null), `error_message` (str|null)

**Research mode**
- **`--research <topic>`** — given a research topic, auto-expands into 5 multi-angle queries (overview/profile/background/works/review) and runs them in sequence. Results include `research_topic` and `research_queries` metadata so AI agents can structure their final report by angle
  - **Cross-angle merge & dedup** — JSON output includes a top-level `merged_results` field: all per-angle results are combined, deduplicated, and sorted. Brief/urls formats append a `[MERGED]` section after the per-angle blocks
  - **Bilingual suffixes** — detects whether the topic contains CJK characters. Chinese topics use Chinese suffixes (overview uses empty suffix = topic itself, then 简介/经历/作品/评价); English topics use English suffixes (overview uses empty suffix, then profile/background/works/reviews)
  - **`--research-angles`** — custom angles overriding the default 5 (e.g. `"overview,profile,timeline,controversy"`); each angle appended directly as a query suffix
- Mutually exclusive with `--query` and `--queries-file`. Supports all output formats. Deterministic expansion (no AI judgment) — same topic always produces same queries

**Observability**
- **Structured logging** (`--log-format {text,json}`, default text) — `json` outputs one JSON object per line `{ts, level, logger, msg, request_id}` for programmatic parsing
- **Request ID propagation** — each run auto-generates an 8-char hex `request_id` threaded through all logs and progress events
- **`--dry-run`** — preview mode: no HTTP requests sent; prints `{dry_run, instances, headers_count, action, ...}` JSON to stdout (`instances` is a list of instance URLs)
- **`--progress`** — JSON Lines events to stderr with `request_id` for real-time tracking

**Engineering**
- Shared `common.py` module — unified retry policy/charset/auth resolution/logging/UA-pool/browser-headers/backoff/error classification/error codes/recovery hints/Wayback fallback/hard-blocked domains/similarity dedup/UTF-8 enforcement/proxy/Retry-After parsing across both scripts
- `search.py --fetch` reuses `fetch.py`'s higher-quality text extractor (no code duplication)
- Structured logging (`--verbose` / `--quiet`) — three levels: default INFO, `--verbose` DEBUG, `--quiet` WARNING. All log output to stderr; stdout reserved for data
- UTF-8 stdout enforcement (`force_utf8_stdout()`) — Windows Python defaults to GBK and crashes on non-ASCII chars; both scripts force UTF-8 at startup
- Windows config discovery — also checks `%APPDATA%/searxng-cli/` in addition to `~/.config/searxng-cli/`
- Fetch failure diagnostics — error output includes `status_code=`, `cause=`, `url=` fields for programmatic distinction of 404 vs 403 vs DNS failure
- Engine/category whitespace normalization (`"google, bing"` → `"google,bing"`)

**Scripts + shared module:**
1. `search.py` — execute searches against a user-supplied instance, with multi-instance failover + retry + auto-fetch + caching + batch + domain filtering
2. `fetch.py` — download and extract readable text or markdown from web pages
3. `common.py` — shared utilities (auth resolution + headers, charset detection, retry policy, UA pool, browser headers, backoff, logging, progress events, error classification + codes, recovery hints, Wayback fallback, hard-blocked domains, similarity dedup, UTF-8 enforcement, proxy, Retry-After parsing) used by both scripts
4. `cache.py` — SQLite-backed result cache (SHA-256 key, TTL, WAL mode, size cap + LRU eviction, schema migration, module-level + class API)
5. `_config.py` — package constants (version, schema version, default user agent, UA pool)

## Default settings

`search.py` ships with opinionated defaults tuned for AI research:

| Setting | Default | Flag to override |
|---------|---------|------------------|
| Instance | **required** — via `-i`, `SEARXNG_INSTANCE` env var, or config file | `-i / --instance` |
| Safe search | **0 (off)** | `-s / --safesearch {0,1,2}` |
| Time range | **year** | `-t / --time-range {day,week,month,year,none}` (none = disabled) |
| Output format | **json** | `-f / --format {json,brief,urls,csv}` |
| Engines | **google,bing,brave,duckduckgo,startpage,wikipedia,wikidata** | `--engines <list>` |

## Quick Start

```bash
# Prerequisites: Python 3.8+
# Optional but recommended:
pip install requests beautifulsoup4

# 1. Search against YOUR instance (instance URL is required)
python scripts/search.py -q "python asyncio tutorial" -i https://my-searxng.example.com

# 2. Multiple instances for failover (comma-separated)
python scripts/search.py -q "rust memory safety" \
  -i https://a.example.com,https://b.example.com --format brief

# 3. Search + auto-fetch top 3 result pages in one command
python scripts/search.py -q "climate policy" -i https://my-searxng.example.com --fetch 3

# 4. Fetch a result page
python scripts/fetch.py -u "https://example.com" --extract text

# 5. Skip -i by configuring the instance once (env var, current shell)
export SEARXNG_INSTANCE="https://my-searxng.example.com,https://backup.example.com"
python scripts/search.py -q "python asyncio tutorial"          # -i not needed

# 6. Or use a config file (./searxng.toml or ~/.config/searxng-cli/searxng.toml
#    or %APPDATA%/searxng-cli/searxng.toml on Windows)
#    [searxng]
#    instance = "https://my-searxng.example.com"
#    # or: instances = ["https://a.example.com", "https://b.example.com"]
#    # Auth (optional, for private instances):
#    auth_basic = "user:password"        # Basic auth
#    auth_bearer = "sk-token-123"        # Bearer token (auth_bearer wins if both set)
#    # Any flag below can also be pre-set here (engines, categories, language,
#    # safesearch, time_range, method, format, timeout, max_retries, proxy,
#    # cache_ttl, cache_max_size, fetch, fetch_timeout, fetch_retries, max_size).
#    # Explicit CLI flags always override config values.
#    # Auth priority: --auth-* > --auth-*-file > searxng.toml > env var
#    Plain list also works in ./instances.txt (one URL per line, # for comments)

# 7. Cache results for 30 minutes (identical queries skip the network)
python scripts/search.py -q "python asyncio" -i https://s.example.com --cache-ttl 30

# 7b. Sort by date (newest first) or disable dedup for raw engine output
python scripts/search.py -q "ai news" -i https://s.example.com --sort-by date --no-dedup

# 8. Batch: run queries from a file (one per line; blank/# lines skipped)
python scripts/search.py --queries-file queries.txt -i https://s.example.com --format json > batch.json

# 9. Domain allowlist + blocklist (applied after search)
python scripts/search.py -q "rust async" -i https://s.example.com \
  --include-domain doc.rust-lang.org,wikipedia.org --exclude-domain pinterest.com

# 10. Route through a corporate proxy (applies to search and fetch)
python scripts/search.py -q "ai news" -i https://s.example.com --proxy http://corp-proxy:8080

# 11. Auth from a file (avoids leaking tokens in shell history)
python scripts/search.py -q "test" -i https://private.example.com --auth-bearer-file ~/.searxng_token

# 12. Cache management (no search performed)
python scripts/search.py --cache-stats        # entries, oldest/newest, size, total, max, evicted, utilization
python scripts/search.py --clear-cache        # delete all entries

# 13. Export results as CSV (great for spreadsheets / data analysis)
python scripts/search.py -q "rust async" -i https://s.example.com --format csv > results.csv

# 14. Use a specific config file (overrides auto-discovered searxng.toml)
python scripts/search.py --config ./my-config.toml -q "test"

# 15. Control log verbosity on stderr
python scripts/search.py -q "test" -i https://s.example.com --verbose   # debug detail
python scripts/search.py -q "test" -i https://s.example.com --quiet     # errors only
```

**Dependency levels:**
| Level | Scripts | What you get |
|-------|---------|-------------|
| Zero deps (stdlib only) | `search.py`, `_config.py` | Full search + auto-fetch |
| `pip install requests` | `fetch.py` | Better HTTP (session reuse, redirect handling) |
| `pip install beautifulsoup4` | `fetch.py` | Higher-quality text extraction |

## When to Use

- **Web search without API keys** — programmatic search results against your own SearXNG instance
- **Privacy-conscious research** — queries routed through your instance, not ad-tech infrastructure
- **Scraping search results** — batch query multiple terms and collect structured results as JSON
- **Fetching search result pages** — follow links from search results and extract clean text

**Don't use for:**
- High-frequency production search without a properly-scaled instance — respect your instance's rate limits
- Guaranteed uptime/accuracy — depends entirely on the instance you supply

## AI Agent Integration Guide

This section documents the structured interfaces that AI agents can rely on
for programmatic integration. All features are designed to be machine-readable
and machine-actionable.

### Output Channels

| Channel | Content | Description |
|---------|---------|-------------|
| stdout | Data | JSON/CSV/text — the only source AI should parse |
| stderr | Logs + Progress | Human-readable logs (default) or JSON Lines events (`--progress`) |
| exit 0 | Success | Results available on stdout |
| exit 1 | Fatal error | Error JSON on stdout (in `--format json` mode) or stderr |
| exit 2 | Empty results | Search succeeded but returned no results |

### Error Code System

In `--format json` mode, errors are emitted as structured JSON on stdout:

```json
{
  "error": "All 2 instances failed. Last error: connection refused",
  "error_code": "E_NETWORK",
  "recovery_hint": "Retry with backoff, or try a different SearXNG instance. Check network connectivity, proxy settings, and instance uptime.",
  "exit_code": 1,
  "query": "search term"
}
```

AI agents can use `error_code` to programmatically decide recovery strategy, and `recovery_hint` for a ready-to-use actionable suggestion:

| Code | Meaning | recovery_hint (abridged) |
|------|---------|--------------------------|
| `E_CONFIG` | Configuration error (no instance resolved) | Provide `-i`/`SEARXNG_INSTANCE`/config file |
| `E_AUTH` | Authentication failed (401/403) | Verify credentials, check token expiry & permissions |
| `E_NETWORK` | Network error (connection refused, timeout, 5xx, all instances failed) | Retry with backoff, switch instance, check proxy |
| `E_RATE_LIMIT` | Rate limited (429) | Wait and retry, reduce frequency, distribute load |
| `E_PARSE` | Parse error (JSON/HTML parsing failed) | Try different instance, switch `--method` |
| `E_EMPTY` | Empty results (exit code 2) | Refine query, broaden `--time-range`, add `--categories` |
| `E_INPUT` | Input error (bad parameters, file not found) | Check query syntax, flag combinations, file paths |
| `E_INTERNAL` | Internal error (unexpected exception) | Re-run with `--verbose`, report bug |
| `E_UNSUPPORTED_MEDIA` | Unsupported binary media type (PDF/docx/xlsx parse failure or unsupported content type) | No auto hint — try a different URL, or install `pdftotext` for PDF parsing |

### JSON Lines Streaming (`--stream`)

For large result sets, `--stream` outputs results as JSON Lines (one JSON
object per line) to stdout, allowing AI agents to process results
incrementally without waiting for the full response:

```bash
python scripts/search.py -q "large topic" -i https://your-instance --stream
```

Output format (each line is a separate JSON object):

```
{"type": "result", "result": {"title": "...", "url": "...", "content": "..."}}
{"type": "result", "result": {"title": "...", "url": "...", "content": "..."}}
{"type": "done", "schema_version": "1.0", "count": 2, "query": "large topic"}
```

- `type: "result"` — one per search result, emitted as soon as available
- `type: "done"` — terminal event with total count and `schema_version`, always emitted last
- `type: "error"` — emitted when the search fails, includes `error_code` and `recovery_hint`:
  ```
  {"type": "error", "error": "All 2 instances failed. Last error: HTTP Error 403", "error_code": "E_AUTH", "recovery_hint": "Verify credentials...", "query": "..."}
  ```
- Exit code 0 on success, 1 on error, 2 on empty results (done event still emitted)

Only valid with `--format json` (single query mode). Using `--stream` with
`--queries-file`, `--research`, or non-json formats raises `E_INPUT` immediately.

### Progress Events (`--progress`)

For long-running operations, `--progress` emits structured JSON Lines events
to stderr, enabling AI agents to track execution progress in real time.
Every event includes a `request_id` field (8-char hex, auto-generated per run)
for correlating all events from a single execution.

```bash
python scripts/search.py -q "research topic" -i https://your-instance --progress --fetch 3
```

Event types (each on its own line, JSON Lines format on stderr):

```jsonl
{"event": "start", "query": "research topic", "instances": 2, "request_id": "a1b2c3d4"}
{"event": "instance_try", "url": "https://instance1.example.com", "attempt": 1, "request_id": "a1b2c3d4"}
{"event": "instance_ok", "url": "https://instance1.example.com", "latency": 0.342, "results": 10, "request_id": "a1b2c3d4"}
{"event": "instance_fail", "url": "https://instance2.example.com", "error": "HTTP 503", "error_code": "E_NETWORK", "request_id": "a1b2c3d4"}
{"event": "cache_hit", "query": "research topic", "ttl": 30, "request_id": "a1b2c3d4"}
{"event": "cache_store", "query": "research topic", "ttl": 30, "request_id": "a1b2c3d4"}
{"event": "page_ok", "pageno": 1, "results": 10, "request_id": "a1b2c3d4"}
{"event": "page_fail", "pageno": 2, "error": "HTTP 503", "request_id": "a1b2c3d4"}
{"event": "fetch_start", "count": 3, "request_id": "a1b2c3d4"}
{"event": "fetch_ok", "url": "https://example.com/page", "chars": 12345, "request_id": "a1b2c3d4"}
{"event": "fetch_fail", "url": "https://bad.example.com", "error": "HTTP 503", "request_id": "a1b2c3d4"}
{"event": "done", "results": 10, "query": "research topic", "request_id": "a1b2c3d4"}
{"event": "error", "error": "connection refused", "error_code": "E_NETWORK", "query": "...", "request_id": "a1b2c3d4"}
```

AI agents can parse these events to:
- Show progress indicators to users
- Detect cache hits (skip waiting)
- Monitor fetch failures and retry strategies
- Correlate errors with specific queries in batch mode
- Track per-page progress via `page_ok`/`page_fail` events
- Correlate all events of a single run via `request_id`

`--progress` and `--verbose` can be used together (progress events on stderr,
debug logs also on stderr). `--progress` events are JSON Lines; `--verbose`
logs are human-readable text.

### Output JSON Schema

The default `--format json` output includes a `schema_version` field so AI
agents can detect breaking changes. Run `--dump-schema` to get the full JSON
Schema document programmatically:

```bash
python scripts/search.py --dump-schema
```

Single query output shape:

```json
{
  "schema_version": "1.0",
  "query": "search term",
  "number_of_results": 10,
  "results": [
    {
      "title": "Result title",
      "url": "https://example.com/page",
      "content": "Snippet text...",
      "engine": "google",
      "score": 1.0,
      "category": "general",
      "published_date": "2024-01-15T10:30:00"
    }
  ],
  "answers": ["Direct answer if available"],
  "corrections": [],
  "suggestions": ["related suggestion"],
  "infoboxes": [],
  "unresponsive_engines": [["engine_name", "error reason"]],
  "fetched": [
    {
      "url": "https://example.com/page",
      "status": "ok",
      "text": "Extracted page content...",
      "text_length": 12345,
      "truncated": false,
      "final_url": "https://example.com/final",
      "user_agent_used": "Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/140.0.0.0 Safari/537.36",
      "anti_bot_detected": false,
      "waf_type": null,
      "fallback_used": null,
      "error_code": null
    }
  ],
  "fetched_source": "json"
}
```

Batch mode (`--queries-file`) wraps results in a unified schema:

```json
{
  "schema_version": "1.0",
  "queries": [
    {"query": "term1", "status": "ok", "results": {...}},
    {"query": "term2", "status": "error", "error": "...", "error_code": "E_NETWORK"}
  ]
}
```

Each batch entry has a `status` field (`"ok"` or `"error"`). Successful entries
contain `results`; failed entries contain `error` and `error_code`.

Fields marked as optional may be absent. The `fetched` and `fetched_source`
fields only appear when `--fetch N` is used.

## Cross-Agent Compatibility

These scripts are **agent-agnostic** — they work with any AI agent that can invoke terminal commands, or directly from a terminal.

**Key design decisions for universal compatibility:**
- Zero external dependencies (stdlib-only for `search.py`)
- Scripts inject their own directory into `sys.path`, so they run from **any** working directory
- Stdout carries data (JSON/text), stderr carries progress/warnings
- Exit codes: 0=success, 1=fatal error, 2=no results/empty
- NO agent-specific API calls or tool dependencies — purely CLI-based, portable across all agent platforms

## Scripts

All scripts live in `scripts/`; run with `python scripts/<name>.py` from any directory. They import `_config.py` for shared constants and self-inject their own directory into `sys.path`.

### 1. `search.py` — Execute SearXNG Search

```
usage: search.py [-h] [--query QUERY] [--instance URL]
                 [--categories CATS] [--language LANG] [--pageno N] [--pages N]
                 [--time-range {day,week,month,year,none}] [--safesearch {0,1,2}]
                 [--engines E] [--method {GET,POST}] [--max-results N]
                 [--format {json,brief,urls,csv}] [--snippet-len N] [--stream]
                 [--progress] [--fetch N] [--fetch-timeout SEC] [--fetch-retries N]
                 [--fetch-report [json]] [--no-fallback] [--referer URL]
                 [--request-delay SEC] [--max-size BYTES] [--output FILE]
                 [--timeout SEC] [--retry N] [--fail-fast] [--serial]
                 [--verify] [--auth-bearer TOKEN] [--auth-bearer-file FILE]
                 [--auth-basic USER:PASS] [--auth-basic-file FILE]
                 [--proxy URL] [--include-domain DOMAINS]
                 [--exclude-domain DOMAINS] [--queries-file FILE]
                 [--research TOPIC] [--research-angles ANGLES]
                 [--cache-ttl MINUTES] [--cache-max-size MB] [--clear-cache]
                 [--cache-stats] [--sort-by {score,date,engine,none}] [--no-dedup]
                 [--similarity-dedup] [--similarity-threshold FLOAT]
                 [--throttle-failure-threshold N] [--throttle-pause-seconds SEC]
                 [--throttle-max-delay SEC] [--log-format {text,json}]
                 [--save-config FILE] [--dry-run]
                 [--config FILE] [--dump-schema] [--verbose] [--quiet] [--version]
```

**What it does:**
1. Takes a search query and resolves one or more instance URLs (`-i`, `SEARXNG_INSTANCE`, or config file — comma-separated for failover)
2. **Multi-instance failover:** if an instance fails (429/5xx/timeout/captcha), automatically tries the next one
3. **Exponential backoff:** retries each instance up to 3 times with jitter on transient errors (403, 429, 502, 503, 504, and connection errors)
4. **Parallel probing (default for 2+ instances):** queries every instance concurrently and returns the first *successful* result in your original instance order — this keeps output deterministic while drastically speeding up failover when an early instance is down/slow. Use `--serial` to disable.
5. Calls the SearXNG API (GET or POST) with `format=json`
6. Falls back to HTML scraping if JSON is blocked
7. HTML parser extracts results + suggestions + answers + infoboxes (full content, never truncated)
8. **Auto-fetch:** `--fetch 3` concurrently downloads top 3 result pages with retry, full browser fingerprint headers, 15-UA deterministic per-domain pool, Retry-After compliance, WAF fingerprint detection, Wayback Machine fallback for 404/403/timeout, adaptive throttling, and `fetch.py`'s readability-lite extractor. `--fetch-report` prints a structured per-URL report to stderr
9. **Health-check mode:** `--verify` probes each instance (reachability / JSON-API support / latency / POST support / engine list / auth status) and prints a report, then exits without searching — use it to validate your instance list
10. **Result caching:** `--cache-ttl 30` stores results for 30 min; identical queries within the TTL skip the network entirely. Cache lives at `$SEARXNG_CACHE_DIR` or `~/.cache/searxng-cli/cache.db` (SQLite, WAL mode, LRU eviction). `--clear-cache` / `--cache-stats` manage it without searching
11. **Batch mode:** `--queries-file FILE` reads one query per line (blank/`#` lines skipped) and runs them in sequence; JSON output is `{"schema_version": "1.0", "queries": [{"query":..., "status": "ok"|"error", ...}]}` (or one block per query in brief/urls). A failed query is recorded but does not abort the batch. Exit codes: 0 if any query returned results, 1 if all errored, 2 if all empty
12. **Dedup + sort + domain filter:** After search (and cache), duplicate URLs are collapsed (default; `--no-dedup` disables), results are sorted (`--sort-by`; default: score descending), and then `--include-domain`/`--exclude-domain` filter by domain. Matching is case-insensitive and ignores a leading `www.`; when a domain is in both lists, exclude wins
13. **Proxy & auth:** `--proxy URL` routes both search and fetch through a proxy; `--auth-bearer` / `--auth-basic` (plus `*-file` variants, `searxng.toml` `auth_basic`/`auth_bearer` fields, and `SEARXNG_BEARER_TOKEN` / `SEARXNG_BASIC_AUTH` env vars) supply credentials. Priority: CLI flag > file > config file > env var
14. **Config defaults:** `searxng.toml` may pre-set most flags; explicit CLI flags always win
15. **Structured errors:** in `--format json` mode, failures print a JSON object `{"error": "...", "error_code": "E_*", "recovery_hint": "...", "exit_code": N, "query": "..."}` to stdout so agents can parse them and decide recovery strategy

**Key options:**
- `--query "your search"` — **required unless** `--verify`, `--queries-file`, `--research`, `--clear-cache`, `--cache-stats`, `--save-config`, or `--dump-schema` is used
- `--instance https://searx.example.org` — **required unless** `SEARXNG_INSTANCE` env var or a config file supplies it; comma-separated list enables failover
- `--queries-file FILE` — read queries from a file (one per line; blank/`#` skipped) and run them in sequence; overrides `--query`
- `--engines google,duckduckgo` — restrict to specific search engines (whitespace around commas is auto-stripped; see default list above)
- `--method POST` — use POST instead of GET (better for long queries)
- `--categories general,news` — comma-separated categories (whitespace auto-stripped)
- `--language zh-CN` — language filter
- `--pageno 1` — page number
- `--pages N` — fetch N pages of results in one run and merge with cross-page dedup; each page cached independently (cache key includes `pageno`)
- `--time-range {day,week,month,year,none}` — time filter (default: `year`; `none` disables time filtering)
- `--safesearch {0,1,2}` — safe search (default: `0` = off)
- `--max-results N` — limit number of results (applied AFTER dedup+sort, so the highest-scoring/newest items are kept)
- `--sort-by {score,date,engine,none}` — sort results (default: `score` descending; `none` preserves instance order). Applied after dedup, before `--max-results`. HTML-fallback results have no score and keep their order
- `--no-dedup` — disable cross-engine deduplication (by default, duplicate URLs — same page ignoring tracking params/fragment — are collapsed, keeping the first occurrence's engine/score)
- `--similarity-dedup` — enable title-based similarity deduplication (SimHash + Jaccard); off by default; auto-skipped when results > 500
- `--similarity-threshold FLOAT` — similarity threshold for `--similarity-dedup` (default: 0.85; higher = stricter)
- `--config FILE` — path to a `searxng.toml` config file; overrides the default auto-discovery (`./searxng.toml` → `~/.config/searxng-cli/searxng.toml` → `%APPDATA%/searxng-cli/searxng.toml` on Windows). Must be the first flag so its values can set defaults for other flags
- `--verbose` / `-v` — show debug-level diagnostics on stderr (HTTP request URLs, response codes, cache keys, retry detail)
- `--quiet` — suppress progress messages and retry notices on stderr; only warnings and errors are shown (no short flag: `-q` is `--query`)
- `--include-domain a.com,b.org` — allowlist; only results from these domains are kept (applied after search)
- `--exclude-domain pinterest.com` — blocklist; results from these domains are dropped (applied after search; wins over include on conflict)
- `--serial` — disable parallel multi-instance probing; search instances strictly one at a time
- `--verify` — health-check mode: verify instances and exit (no search); combine with `--format brief` for a table or `--format json` for machine-readable output
- `--format json` — full JSON (default); includes `fetched` array when `--fetch` is used; errors are emitted as JSON to stdout
- `--format brief` — title + URL + full snippet (no truncation by default; use `--snippet-len 200` to cap)
- `--format urls` — only result URLs
- `--format csv` — CSV export (title,url,engine,score,published_date,content); in batch mode (`--queries-file`), all queries merge into one CSV with a `query` column
- `--fetch N` — after search, auto-fetch full text of top N result pages (concurrent, stdlib only)
- `--fetch-timeout 10` — timeout per page fetch (default: 10s)
- `--fetch-retries 3` — max retries per page fetch
- `--max-size BYTES` — cap page size (default: unlimited; e.g. `5242880` for 5MB)
- `--retry 5` — max retries per instance (default: 3)
- `--timeout 15` — request timeout in seconds
- `--fail-fast` — use only the first instance, don't fail over to the rest
- `--proxy URL` — HTTP/HTTPS proxy for both search and fetch (e.g. `http://corp-proxy:8080`)
- `--cache-ttl MINUTES` — cache results for N minutes (default: `0` = disabled); identical queries within the TTL skip the network
- `--cache-max-size MB` — cache size cap in MB with LRU eviction (default: `0` = defer to `$SEARXNG_CACHE_MAX_SIZE_BYTES` env var or 100MB built-in default; when cap exceeded, least-recently-accessed entries evicted; set a very large value for effectively unlimited)
- `--clear-cache` — delete all cached entries and exit (no search)
- `--cache-stats` — print cache statistics (entries, oldest_created_at, newest_created_at, path, size_bytes, total_bytes, max_size_bytes, evicted_count, utilization_pct) and exit
- `--dump-schema` — print the JSON Schema for `--format json` output and exit; lets AI agents programmatically discover field names and types without parsing prose docs
- `--auth-bearer TOKEN` — `Authorization: Bearer` header for private instances
- `--auth-bearer-file FILE` — read Bearer token from a file (first non-empty, non-`#` line); also honors `SEARXNG_BEARER_TOKEN` env var
- `--auth-basic USER:PASS` — `Authorization: Basic` header (auto base64-encoded)
- `--auth-basic-file FILE` — read `user:pass` from a file (first non-empty, non-`#` line); also honors `SEARXNG_BASIC_AUTH` env var
- `--log-format {text,json}` — structured logging format (default: `text`); `json` outputs one JSON object per line `{ts, level, logger, msg, request_id}`
- `--dry-run` — preview mode: no HTTP requests; prints `{dry_run, instances, headers_count, action, ...}` JSON to stdout (`instances` is a list); supports search/research/batch/verify
- `--throttle-failure-threshold N` — consecutive failures before doubling delay + halving concurrency (default: 3)
- `--throttle-pause-seconds SEC` — global pause on 429 (default: 30)
- `--throttle-max-delay SEC` — max adaptive delay (default: 10)
- `--research-angles "a,b,c"` — custom research angles overriding the default 5; each angle appended directly as a query suffix
- `--save-config FILE` — save current CLI args as a `searxng.toml` config file and exit
- `--fetch-report [json]` — `--fetch-report json` outputs a full JSON report (items + summary); original `--fetch-report` (no arg) keeps text format
- `--version` — print version and exit

**Completion criterion:** Outputs valid JSON with `results` array. Non-zero exit on total failure (all instances exhausted).

### 2. `fetch.py` — Fetch & Extract Web Page Content

```
usage: fetch.py [-h] --url URL [--extract {text,html,markdown}]
                [--timeout SEC] [--retries N] [--max-size BYTES]
                [--user-agent STR] [--encoding CHARSET]
                [--no-redirect] [--no-fallback] [--referer URL] [--proxy URL]
                [--output FILE] [--auth-bearer TOKEN] [--auth-bearer-file FILE]
                [--auth-basic USER:PASS] [--auth-basic-file FILE]
                [--verbose] [--quiet] [--version]
```

**What it does:**
1. Downloads a web page via HTTP GET with retry + exponential backoff
2. Retries on 429/5xx/connection errors (3x default) with Retry-After header compliance and capped 60s backoff; falls back through the 15-UA deterministic pool with full browser fingerprint headers; reuses a requests.Session for connection pooling + cookie persistence. **Both paths handle Content-Encoding decompression**: stdlib urllib path handles gzip/deflate/br (urllib doesn't auto-decompress any of them); requests path handles br manually (requests auto-decompresses gzip/deflate but not Brotli unless the brotli package is installed)
3. **No size limit by default** — full page content returned; use `--max-size` for a cap
4. Detects charset from HTTP headers, HTML meta tags, or UTF-8 fallback
5. Extracts readable content using tree-based parsers (stdlib or BeautifulSoup); readability-lite text-density fallback when `<article>`/`<main>`/`role="main"`/content-class `<div>` are all missing and only `<body>` remains
6. Outputs clean text, raw HTML, or properly-converted Markdown (with correct nested-link handling)
7. **PDF/document parsing** (regardless of `--extract` mode): PDF via `pdftotext` subprocess, `.docx`/`.xlsx` via stdlib `zipfile`. Unsupported binary types return `E_UNSUPPORTED_MEDIA`

**Key options:**
- `--url https://...` — required
- `--extract text` — clean readable text (default); PDF/docx/xlsx parsed automatically regardless of extract mode (see "What it does" #7)
- `--extract html` — raw HTML
- `--extract markdown` — Markdown conversion (tree-based; handles nested tags, GFM tables, fenced code blocks, blockquotes, inline code, ordered/unordered/nested lists, definition lists, images, emphasis)
- `--encoding gbk` — force charset for non-UTF-8 pages (else auto-detected from HTTP header / HTML meta / UTF-8 fallback)
- `--timeout 15` — request timeout in seconds (split into connect/read tuple internally)
- `--retries 3` — max retries on transient errors (429/5xx/connection); falls back through the 15-UA pool when blocked
- `--max-size BYTES` — cap page size (default: unlimited; e.g. `5242880` for 5MB)
- `--user-agent STR` — custom User-Agent header (overrides per-domain UA selection)
- `--no-redirect` — do **not** follow HTTP 3xx redirects (implemented for both the requests and stdlib paths)
- `--no-fallback` — disable Wayback Machine fallback for 404/403/timeout and anti-bot-blocked pages (hard-blocked domains like baike.baidu.com, zhihu.com still prefer Wayback)
- `--referer URL` — set Referer header to disguise traffic source
- `--proxy URL` — HTTP/HTTPS proxy (e.g. `http://corp-proxy:8080`); respects existing `HTTP_PROXY`/`HTTPS_PROXY` env vars when omitted
- `--output FILE` — save to file instead of stdout
- `--auth-bearer TOKEN` / `--auth-bearer-file FILE` — `Authorization: Bearer` header; file variant reads first non-empty, non-`#` line; also honors `SEARXNG_BEARER_TOKEN` env var
- `--auth-basic USER:PASS` / `--auth-basic-file FILE` — `Authorization: Basic` header (auto base64-encoded); file variant + `SEARXNG_BASIC_AUTH` env var also supported
- `--verbose` / `-v` — show debug-level diagnostics on stderr
- `--quiet` — suppress progress messages on stderr; only warnings and errors are shown
- `--version` — print version and exit

**Extraction strategy (text mode):**
1. Strip non-content elements (script, style, nav, footer, header)
2. Extract main content in priority order: `<article>` → `<main>` → `role="main"` → content-class `<div>` (class~=`content`/`article`/`post`/`entry`) → `<body>`
3. If only `<body>` matched (all higher-priority elements missing), run readability-lite to pick the highest text-density `<div>`/`<section>`/`<article>` node (scored by text density + `<p>` count weighting; filters out nav/sidebar/footer by class/id)
4. Collapse whitespace, output clean UTF-8
5. Warn if extracted text < 500 chars (likely JS-heavy or bot-blocked)

**Completion criterion:** Outputs page content. Non-zero exit on HTTP failure. Stderr carries warnings for low-confidence extraction.