Keep runtime activation tokens out of the sites database (#2973)

* Keep runtime activation tokens out of the sites database

* Document the activation token cache

docs/source/locale/zh_CN/LC_MESSAGES/settings.po

@@ -440,3 +440,60 @@ msgstr ""
 #~ "``settings.json`` 中的 ``cloudflare_bypass`` 配置块用于配置"
 #~ " :ref:`cloudflare-bypass` 所述的可选绕过机制。默认值如下:"
 
+
+#: ../../source/settings.rst:107
+msgid "Activation token cache"
+msgstr "激活令牌缓存"
+
+#: ../../source/settings.rst:110
+msgid ""
+"A handful of sites reject anonymous requests until a short-lived token is "
+"obtained: Twitter needs a guest token, Vimeo a JWT, Weibo and Wikimapia a "
+"session cookie, OnlyFans a signed header pair, ProtonMail a bearer token. "
+"Maigret mints these during a run, the first time such a site answers with a "
+"challenge instead of a profile page."
+msgstr ""
+"少数站点会拒绝匿名请求,必须先获取一个短期令牌:Twitter 需要 guest token,Vimeo 需要 "
+"JWT,Weibo 和 Wikimapia 需要会话 cookie,OnlyFans 需要一对签名请求头,ProtonMail "
+"需要 bearer token。当这类站点返回的是验证页面而非个人主页时,Maigret 会在本次运行中即时获取这些令牌。"
+
+#: ../../source/settings.rst:112
+msgid "Minted values are cached in:"
+msgstr "获取到的值会缓存在:"
+
+#: ../../source/settings.rst:118
+msgid ""
+"so the next run starts with a warm token and skips the extra round trip. The"
+" file is created with ``0600`` permissions because these tokens are session "
+"credentials — treat it as one, and do not attach it to a bug report."
+msgstr ""
+"这样下次运行即可直接使用仍然有效的令牌,省去一次额外的往返请求。该文件以 ``0600`` "
+"权限创建,因为这些令牌属于会话凭据——请按凭据对待,不要将其附在问题报告中。"
+
+#: ../../source/settings.rst:120
+msgid ""
+"**What is stored:** only the header values that differ from what the site "
+"database shipped. Storing the whole header set would pin the shipped values,"
+" so a later database update could never change them again."
+msgstr ""
+"**缓存的内容:** 仅保存与站点数据库自带值不同的请求头。若把整组请求头都存下来,就会把自带值固定住,"
+"以后的数据库更新将再也无法修改它们。"
+
+#: ../../source/settings.rst:122
+msgid ""
+"**Safe to delete:** yes. The next run mints whatever it needs again; the "
+"only cost is one extra request per affected site."
+msgstr ""
+"**可以安全删除:** 可以。下次运行会重新获取所需的令牌,代价仅仅是每个相关站点多一次请求。"
+
+#: ../../source/settings.rst:124
+msgid ""
+"**Why it is not kept in the site database:** the database is package data. "
+"It is versioned, checksum-verified and replaced wholesale by :ref:`database-"
+"auto-update`, and on a system-wide install (distribution package, snap, "
+"``/nix/store``) its directory is not writable at all. Tokens are per-user "
+"state with a different lifetime, so they live in a separate file."
+msgstr ""
+"**为什么不存放在站点数据库中:** 数据库属于软件包自带数据,它有版本、会做校验和验证,并且会被 "
+":ref:`database-auto-update` 整体替换;而在系统级安装(发行版软件包、snap、``/nix/store``)下,"
+"它所在的目录根本不可写。令牌是每个用户各自的状态,生命周期也不同,因此单独存放。"

docs/source/settings.rst

@@ -102,6 +102,27 @@ This is recommended for **Docker containers**, **CI pipelines**, and **air-gappe
 
 **Using a custom database** with ``--db`` always skips auto-update — you are explicitly choosing your data source.
 
+.. _activation-token-cache:
+
+Activation token cache
+----------------------
+
+A handful of sites reject anonymous requests until a short-lived token is obtained: Twitter needs a guest token, Vimeo a JWT, Weibo and Wikimapia a session cookie, OnlyFans a signed header pair, ProtonMail a bearer token. Maigret mints these during a run, the first time such a site answers with a challenge instead of a profile page.
+
+Minted values are cached in:
+
+.. code-block:: console
+
+   ~/.maigret/activation.json
+
+so the next run starts with a warm token and skips the extra round trip. The file is created with ``0600`` permissions because these tokens are session credentials — treat it as one, and do not attach it to a bug report.
+
+**What is stored:** only the header values that differ from what the site database shipped. Storing the whole header set would pin the shipped values, so a later database update could never change them again.
+
+**Safe to delete:** yes. The next run mints whatever it needs again; the only cost is one extra request per affected site.
+
+**Why it is not kept in the site database:** the database is package data. It is versioned, checksum-verified and replaced wholesale by :ref:`database-auto-update`, and on a system-wide install (distribution package, snap, ``/nix/store``) its directory is not writable at all. Tokens are per-user state with a different lifetime, so they live in a separate file.
+
 Cloudflare webgate
 ------------------
 

maigret/activation.py

@@ -1,10 +1,15 @@
 import json
+import os
+import os.path as path
 import re
 from http.cookiejar import MozillaCookieJar
 from http.cookies import Morsel
+from typing import Dict
 
 from aiohttp import ClientSession, CookieJar
 
+from .db_updater import MAIGRET_HOME
+
 
 class ParsingActivator:
     @staticmethod
@@ -197,3 +202,73 @@ def import_aiohttp_cookies(cookiestxt_filename):
     cookies.update_cookies(cookies_list)
 
     return cookies
+
+
+ACTIVATION_CACHE_PATH = path.join(MAIGRET_HOME, "activation.json")
+
+
+def _activation_sites(db):
+    return [site for site in db.sites if site.activation]
+
+
+def load_activation_cache(db, logger) -> Dict[str, Dict[str, str]]:
+    """Apply per-user cached activation headers to the sites database.
+
+    Tokens minted at runtime (guest tokens, JWTs, session cookies) are user
+    state, not package data, so they are kept in the user's home instead of
+    being written back into the shipped database — which is read-only on any
+    system-wide install.
+
+    Returns the pre-overlay headers, needed by :func:`save_activation_cache`
+    to tell a minted value from one that simply came with the database.
+    """
+    baseline = {site.name: dict(site.headers) for site in _activation_sites(db)}
+
+    try:
+        with open(ACTIVATION_CACHE_PATH, encoding="utf-8") as f:
+            cache = json.load(f)
+    except FileNotFoundError:
+        return baseline
+    except (OSError, ValueError) as e:
+        logger.debug(f"Ignoring unreadable activation cache: {e}")
+        return baseline
+
+    for site in _activation_sites(db):
+        cached = cache.get(site.name)
+        if isinstance(cached, dict):
+            # Assign rather than update: MaigretSite.headers defaults to a
+            # mutable class attribute shared by every site that doesn't
+            # declare its own, and updating it in place would leak these
+            # tokens onto all of them.
+            site.headers = {**site.headers, **cached}
+
+    return baseline
+
+
+def save_activation_cache(db, baseline: Dict[str, Dict[str, str]], logger) -> None:
+    """Persist only the header values that this run actually minted.
+
+    Storing the full header set would freeze whatever the database shipped at
+    the time, so a later database update could no longer change those headers.
+    """
+    cache = {}
+    for site in _activation_sites(db):
+        was = baseline.get(site.name, {})
+        minted = {k: v for k, v in site.headers.items() if was.get(k) != v}
+        if minted:
+            cache[site.name] = minted
+
+    if not cache:
+        return
+
+    try:
+        os.makedirs(MAIGRET_HOME, exist_ok=True)
+        # These are session credentials (guest tokens, JWTs, cookies), so the
+        # file is created 0600 rather than inheriting the umask.
+        fd = os.open(
+            ACTIVATION_CACHE_PATH, os.O_WRONLY | os.O_CREAT | os.O_TRUNC, 0o600
+        )
+        with os.fdopen(fd, "w", encoding="utf-8") as f:
+            json.dump(cache, f, indent=2, ensure_ascii=False)
+    except OSError as e:
+        logger.debug(f"Could not write the activation cache: {e}")

maigret/maigret.py

@@ -54,6 +54,7 @@ from .report import (
     save_markdown_report,
 )
 from .result import SiteResult
+from .activation import load_activation_cache, save_activation_cache
 from .sites import MaigretDatabase
 from .submit import Submitter
 from .utils import get_dict_ascii_tree, is_plausible_username
@@ -569,6 +570,22 @@ def setup_arguments_parser(settings: Settings):
     return parser
 
 
+def save_db_safely(db: MaigretDatabase, db_file: str, logger) -> bool:
+    """Persist the sites database, tolerating a read-only installation.
+
+    The bundled database lives inside the package, so on a system-wide
+    install (distro package, snap, /nix/store) its directory belongs to
+    root or is mounted read-only. Returns False instead of raising, so a
+    run that already produced its report doesn't die on a cache write.
+    """
+    try:
+        db.save_to_file(db_file)
+        return True
+    except OSError as e:
+        logger.debug(f'Could not write the database to {db_file}: {e}')
+        return False
+
+
 async def main():
     # Logging
     log_level = logging.ERROR
@@ -704,6 +721,9 @@ async def main():
             )
         else:
             raise
+
+    activation_baseline = load_activation_cache(db, logger)
+
     get_top_sites_for_id = lambda x: db.ranked_sites_dict(
         top=args.top_sites,
         tags=args.tags,
@@ -718,8 +738,10 @@ async def main():
     if args.new_site_to_submit:
         submitter = Submitter(db=db, logger=logger, settings=settings, args=args)
         is_submitted = await submitter.dialog(args.new_site_to_submit, args.cookie_file)
-        if is_submitted:
-            db.save_to_file(db_file)
+        if is_submitted and not save_db_safely(db, db_file, logger):
+            query_notify.warning(
+                f'The new site was not saved: {db_file} is not writable'
+            )
         await submitter.close()
 
     # Database self-checking
@@ -754,8 +776,10 @@ async def main():
                 'y',
                 '',
             ):
-                db.save_to_file(db_file)
-                print('Database was successfully updated.')
+                if save_db_safely(db, db_file, logger):
+                    print('Database was successfully updated.')
+                else:
+                    print(f'Database was not updated: {db_file} is not writable.')
             else:
                 print('Updates will be applied only for current search session.')
 
@@ -1112,8 +1136,7 @@ async def main():
             except Exception as e:
                 query_notify.warning(f'AI analysis failed: {e}')
 
-    # update database
-    db.save_to_file(db_file)
+    save_activation_cache(db, activation_baseline, logger)
 
 
 def run():

tests/test_activation.py

@@ -1,6 +1,9 @@
 """Maigret activation test functions"""
 
 import inspect
+import json
+import os
+
 import yarl
 
 import aiohttp
@@ -8,7 +11,14 @@ import pytest
 from unittest.mock import Mock
 
 from tests.conftest import LOCAL_SERVER_PORT
-from maigret.activation import ParsingActivator, import_aiohttp_cookies
+from maigret import activation as activation_cache
+from maigret.activation import (
+    ParsingActivator,
+    import_aiohttp_cookies,
+    load_activation_cache,
+    save_activation_cache,
+)
+from maigret.sites import MaigretDatabase, MaigretSite
 
 COOKIES_TXT = """# HTTP Cookie File downloaded with cookies.txt by Genuinous @genuinous
 # This file can be used by wget, curl, aria2c and other standard compliant tools.
@@ -409,3 +419,139 @@ async def test_wikimapia_activation_no_token_leaves_cookie_untouched():
     await ParsingActivator.wikimapia(site, Mock(), html="<html>no challenge here</html>")
 
     assert site.headers["Cookie"] == "verified=1"
+
+
+@pytest.fixture
+def activation_db(monkeypatch, tmp_path):
+    """Two sites, one that mints tokens at runtime and one that doesn't."""
+    db = MaigretDatabase()
+    db.update_site(
+        MaigretSite(
+            'Twitter',
+            {
+                'url': 'https://twitter.com/{username}',
+                'urlMain': 'https://twitter.com/',
+                'activation': {'method': 'twitter', 'marks': ['nope']},
+                'headers': {'x-guest-token': 'from-db', 'accept': 'text/html'},
+            },
+        )
+    )
+    db.update_site(
+        MaigretSite(
+            'Plain',
+            {
+                'url': 'https://example.com/{username}',
+                'urlMain': 'https://example.com/',
+                'headers': {'accept': 'text/html'},
+            },
+        )
+    )
+    monkeypatch.setattr(activation_cache, 'MAIGRET_HOME', str(tmp_path))
+    monkeypatch.setattr(
+        activation_cache, 'ACTIVATION_CACHE_PATH', str(tmp_path / 'activation.json')
+    )
+    return db
+
+
+def test_activation_cache_persists_only_minted_headers(activation_db):
+    """The database ships baseline headers; only what the run minted is cached.
+
+    Storing the full header set would pin whatever the database shipped, so a
+    later database update could never change those headers again.
+    """
+    logger = Mock()
+    baseline = load_activation_cache(activation_db, logger)
+
+    # simulate what ParsingActivator does mid-run
+    activation_db.sites_dict['Twitter'].headers['x-guest-token'] = 'minted'
+
+    save_activation_cache(activation_db, baseline, logger)
+
+    written = json.loads(open(activation_cache.ACTIVATION_CACHE_PATH).read())
+    assert written == {'Twitter': {'x-guest-token': 'minted'}}
+    # the unchanged baseline header is not copied into the cache
+    assert 'accept' not in written['Twitter']
+
+
+def test_activation_cache_round_trip_applies_token(activation_db):
+    logger = Mock()
+    baseline = load_activation_cache(activation_db, logger)
+    activation_db.sites_dict['Twitter'].headers['x-guest-token'] = 'minted'
+    save_activation_cache(activation_db, baseline, logger)
+
+    # a fresh run loads the database again, with the shipped value
+    activation_db.sites_dict['Twitter'].headers['x-guest-token'] = 'from-db'
+    load_activation_cache(activation_db, logger)
+
+    assert activation_db.sites_dict['Twitter'].headers['x-guest-token'] == 'minted'
+
+
+def test_activation_cache_skips_sites_without_activation(activation_db):
+    logger = Mock()
+    baseline = load_activation_cache(activation_db, logger)
+    activation_db.sites_dict['Plain'].headers['accept'] = 'application/json'
+
+    save_activation_cache(activation_db, baseline, logger)
+
+    assert not os.path.exists(activation_cache.ACTIVATION_CACHE_PATH)
+
+
+def test_activation_cache_ignores_corrupt_file(activation_db):
+    logger = Mock()
+    with open(activation_cache.ACTIVATION_CACHE_PATH, 'w') as f:
+        f.write('{ this is not json')
+
+    baseline = load_activation_cache(activation_db, logger)
+
+    assert baseline['Twitter']['x-guest-token'] == 'from-db'
+    assert activation_db.sites_dict['Twitter'].headers['x-guest-token'] == 'from-db'
+    assert logger.debug.called
+
+
+def test_activation_cache_is_not_world_readable(activation_db):
+    """The cache holds session credentials, so the umask must not decide."""
+    logger = Mock()
+    baseline = load_activation_cache(activation_db, logger)
+    activation_db.sites_dict['Twitter'].headers['x-guest-token'] = 'minted'
+
+    save_activation_cache(activation_db, baseline, logger)
+
+    mode = os.stat(activation_cache.ACTIVATION_CACHE_PATH).st_mode & 0o777
+    assert mode == 0o600, f"expected 0600, got {mode:o}"
+
+
+def test_activation_cache_does_not_leak_into_shared_headers(monkeypatch, tmp_path):
+    """MaigretSite.headers defaults to a mutable class attribute.
+
+    A site with an activation block but no headers of its own shares that
+    object with ~3000 other sites, so an in-place update would attach the
+    cached token to every one of them.
+    """
+    logger = Mock()
+    db = MaigretDatabase()
+    db.update_site(
+        MaigretSite(
+            'NoHeaders',
+            {
+                'url': 'https://a.example/{username}',
+                'urlMain': 'https://a.example/',
+                'activation': {'method': 'twitter', 'marks': ['nope']},
+            },
+        )
+    )
+    db.update_site(
+        MaigretSite(
+            'Bystander',
+            {'url': 'https://b.example/{username}', 'urlMain': 'https://b.example/'},
+        )
+    )
+    cache_file = tmp_path / 'activation.json'
+    cache_file.write_text(json.dumps({'NoHeaders': {'x-guest-token': 'secret'}}))
+    monkeypatch.setattr(activation_cache, 'MAIGRET_HOME', str(tmp_path))
+    monkeypatch.setattr(activation_cache, 'ACTIVATION_CACHE_PATH', str(cache_file))
+
+    load_activation_cache(db, logger)
+
+    assert db.sites_dict['NoHeaders'].headers['x-guest-token'] == 'secret'
+    assert 'x-guest-token' not in db.sites_dict['Bystander'].headers
+    assert 'x-guest-token' not in MaigretSite.headers

tests/test_maigret.py

@@ -1,12 +1,14 @@
 """Maigret main module test functions"""
 
 import asyncio
+import json
+import os
 import copy
 from unittest.mock import Mock, patch
 
 import pytest
 
-from maigret.maigret import self_check, maigret
+from maigret.maigret import self_check, maigret, save_db_safely
 from maigret.maigret import (
     extract_ids_from_page,
     extract_ids_from_results,
@@ -293,3 +295,37 @@ def test_main_entrypoint_handles_top_level_keyboard_interrupt_cleanly():
     # No Python traceback may leak through
     assert "Traceback" not in result.stderr
     assert "KeyboardInterrupt" not in result.stderr
+
+
+@pytest.mark.skipif(os.geteuid() == 0, reason="root ignores file permissions")
+def test_save_db_safely_returns_false_on_readonly_target(default_db, tmp_path):
+    """A read-only install must not take down a run that already finished.
+
+    Distro packages, snaps and /nix/store put the bundled database in a
+    directory the user cannot write, and it is written unconditionally at
+    the end of every run to cache activation tokens.
+    """
+    logger = Mock()
+    readonly_dir = tmp_path / "site-packages"
+    readonly_dir.mkdir()
+    target = readonly_dir / "data.json"
+    target.write_text("{}")
+    readonly_dir.chmod(0o555)
+    target.chmod(0o444)
+
+    try:
+        assert save_db_safely(default_db, str(target), logger) is False
+        assert logger.debug.called
+        # the untouched file is still the one that was there before
+        assert target.read_text() == "{}"
+    finally:
+        readonly_dir.chmod(0o755)
+        target.chmod(0o644)
+
+
+def test_save_db_safely_writes_when_target_is_writable(default_db, tmp_path):
+    logger = Mock()
+    target = tmp_path / "data.json"
+
+    assert save_db_safely(default_db, str(target), logger) is True
+    assert "sites" in json.loads(target.read_text())