Files
WebHooker/docs/zh/guide/storage.md
T
翅膀喵andRhenCloud 978c53eeb3 fix(security): harden webhook delivery pipeline and group storage (#11)
* fix(log): preserve structured error details

Error objects lose their useful details when serialized as plain JSON.
Keep the message, stack and cause in structured log output.

* fix(runtime): bound shared in-flight waits

A canceled request can leave its shared promise pending forever.
Bound joins and coordinate a single replacement so later requests
make progress without multiplying upstream work.

* fix(bot): reject oversized callback bodies

Reject declared oversized bot callbacks before reading their bodies.
This keeps unauthenticated requests from wasting worker resources.

* test(d1): exercise real sqlite transactions

Exercise real foreign keys, transactions and conditional statements
instead of mocks that cannot reproduce lost updates or cascades.

* fix(storage): preserve claims across kv fallback

Only missing tables justify KV fallback; other failures must surface.
Remember fallback writes so migration to D1 cannot forget claims
and replay completed work. Share marker checks without extra writes.

* fix(filter): use isolated linear-time patterns

Backtracking patterns can exhaust worker CPU on untrusted payloads.
Use linear-time matching and separate bounded cache namespaces so
glob and regex entries cannot corrupt each other.

* fix(filter): read each event branch correctly

Deployment and check branches live in event-specific nested fields.
Read those fields and remove only the actual leading ref prefix.

* fix(filter): read only own payload properties

Inherited object properties are not webhook payload data.
Read own properties only so JSONPath cannot traverse prototypes.

* fix(filter): validate executable condition trees

Malformed trees and unsupported patterns must not silently turn an
exclude rule into a pass. Validate the shape and only the patterns
the evaluator actually executes.

* fix(i18n): bound and cache translation loads

Repeated misses and simultaneous override loads spend KV reads for
the same language. Cache confirmed misses and share bounded loads
without caching read failures as missing translations.

* i18n(bot): explain command and account failures

Explain account linking, command failures and skipped delivery
outcomes consistently in both supported languages.

* fix(github): bound and reuse app token fetches

Repeated token exchanges waste requests and JWT work for every event.
Reuse valid installation tokens, share refreshes and back off briefly
after failures while keeping every shared wait bounded.

* fix(github): restrict enrichment request urls

Webhook-provided URLs must never receive an installation token on
another host. Restrict enrichment to GitHub repository API paths,
reject redirects and bound optional requests.

* fix(formatter): preserve invalid code points

Out-of-range numeric entities can throw while formatting a delivery.
Keep invalid code points verbatim instead of aborting the message.

* fix(formatter): preserve bounded markdown text

Code spans, fences, indentation and autolinks are Markdown content,
not HTML tags to strip. Preserve them while bounding cleanup input
and reusing results across routes for the same event.

* fix(formatter): strip only leading ref prefixes

A short branch or tag may itself contain refs/heads or refs/tags.
Strip one leading full-ref prefix at the caller and keep short names
unchanged when generating labels and links.

* fix(gitea): link push authors to their forge

A Gitea author is not the same-named GitHub account.
Build author profile links from the originating forge.

* fix(formatter): render deployment state values

The deployment status object stringifies as [object Object].
Render its actual state in the status field.

* fix(formatter): read security severity levels

The scanner severity is not the security severity of an alert.
Prefer the security level so critical and high alerts get the right
label and color.

* fix(formatter): hide actions on other forges

Action buttons call the GitHub API as the linked user.
Hide them for other forges to avoid operating on a same-named
GitHub repository.

* fix(check): isolate updates by check suite

Same-named checks from different suites must not edit one message.
Include the suite or app identity and retain the former key for
controlled adoption of existing tracked messages.

* fix(logs): scope queries before limiting results

Filter authorized groups in SQL before LIMIT and aggregate only
visible rows. Preserve current group attribution for late writes,
surface read failures and combine related totals in one scan.

* perf(d1): index pruning and failed send queries

Pruning and failure views should not scan all historical rows.
Index the high-value access paths without adding a write-heavy
index for infrequent delivery-detail lookups.

* fix(discord): bound individual api requests

An unbounded API request can outlive an update-chain lock.
Bound each request so another holder cannot race a late edit.

* fix(discord): defer long rate-limit waits

Long rate-limit sleeps block delivery progress inside a worker.
Return a retryable 429 for the queue when the wait exceeds the bound.

* fix(discord): encode channel path segments

Stored channel and message ids are path segments, not URL paths.
Encode them so aliases cannot address another Discord resource.

* fix(discord): reuse nonces after lost responses

A successful create may lose its response. Reuse a stable nonce
with enforce_nonce so retries recover that message instead of
posting duplicates or accepting an unreadable success.

* fix(discord): constrain embeds and role mentions

Respect combined embed and component limits as well as field caps.
Restrict allowed mentions to validated route roles so payload text
cannot notify everyone.

* fix(discord): resume bounded command syncs

Application changes, command rollbacks and large guild lists can
leave commands stale indefinitely. Track the current hash per app
and resume bounded paginated work without needless KV writes.

* fix(github): verify installation management

An installation id is user-controlled, and repository access is not
installation management. Verify the owning account or an active
organization admin using the user token and read:org scope.

* fix(github): resolve immutable member ids

A login can be renamed and reclaimed before its first sign-in.
Resolve the current immutable user id when membership is granted
so a future holder of the login cannot inherit access.

* fix(auth): revoke stale sessions and tokens

Revoked or rotated credentials must stop working despite stale KV
replicas. Track the live token and revocations in D1 and preserve
immutable identity bindings rather than trusting a recycled login.

* fix(auth): bind sessions and tokens to identity

Automatic retries can repeat non-idempotent user actions.
Disable those retries and expose a write hook so bot handlers can
record completion immediately before the GitHub write.

* fix(discord): defer and deduplicate interactions

Acknowledge slow actions before Discord times out and deduplicate
interaction writes. Preserve comment text and reject edits whose
source changed after the modal opened.

* fix(telegram): always verify webhook secrets

Webhook verification must not silently disable itself when the
optional secret is absent. Derive a stable secret from the bot
token and reject callbacks when verification is unavailable.

* fix(telegram): bound individual api requests

Bound individual platform calls so they cannot outlive the lock
protecting a message update or stall a whole delivery invocation.

* fix(telegram): honor nested rate-limit delays

Telegram places retry_after inside parameters, not at the top level.
Read that value and defer long waits to the delivery queue.

* fix(telegram): report unreadable send results

An unreadable successful response may already have sent the message.
Report an explicit uncertain result instead of retrying a create
without an idempotency key or pretending it succeeded.

* fix(telegram): keep editable messages as text

Guessing photo versus text from message content makes later edits
use the wrong API and a smaller caption budget. Always send text
and use the avatar as a link preview.

* fix(telegram): render hard breaks as newlines

Telegram HTML does not support br tags. Render Markdown hard breaks
as plain newlines so otherwise valid messages are accepted.

* fix(telegram): clamp visible html characters

Telegram counts visible characters, not HTML source bytes.
Truncate without splitting tags, entities or surrogate pairs.

* fix(telegram): honor the author emoji toggle

The author marker must obey the group emoji preference just like
formatter-added decoration.

* fix(telegram): keep completed commands claimed

Redelivery after a completed write must never post another comment.
Keep completion claims despite reply failures, scope updates per bot
and only issue account-login links in private chats.

* fix(feishu): bound individual api requests

Bound individual platform calls so they cannot outlive the lock
protecting a message update or stall a whole delivery invocation.

* fix(feishu): classify retryable api failures

Business error codes can be transient even in HTTP 200 responses.
Classify token and rate-limit errors for retry and defer long waits.

* fix(feishu): retry sends with stable uuids

An unreadable response does not prove a send failed.
Reuse the same uuid for retries so recovering the result cannot
create a second Feishu message.

* fix(feishu): refresh and evict tokens safely

Separate app token caches, stop reusing refused tokens and share
refreshes. Cache failures briefly and keep valid tokens in memory
when a best-effort KV cache write fails.

* fix(feishu): use the card update endpoint

Delayed card updates use the card update endpoint; the config
endpoint cannot update an interaction response.

* fix(feishu): neutralize untrusted mentions

Forge text is untrusted and can contain at-all tags.
Neutralize mentions in card bodies and footers before rendering.

* fix(feishu): authenticate and dedupe callbacks

Authenticate signed or token-verified callbacks before dispatching
commands. Decode encrypted envelopes and handle card actions using
the correct fields without claiming irrelevant chat events.

* fix(target): normalize actual destinations

Different target spellings can name the same actual destination.
Normalize channel, thread and topic identities for ownership checks
and per-route duplicate suppression.

* fix(config): transact per-group configuration

Whole-table writes and cached snapshots can overwrite other groups
or resurrect removed configuration. Use per-group guarded writes,
one-time identity migration and explicit unreadable-row failures.

* fix(group): preserve identity through lifecycle

Group names are mutable and reusable, so they cannot own durable
state. Preserve immutable identities across lifecycle operations and
commit invite use with membership under transaction-time guards.

* fix(oauth): confirm account links and installs

A login link or installation id is not proof of account ownership.
Verify installation management, bind browser state and require an
explicit account-link confirmation before granting chat access.

* fix(queue): fence leases and record target outcomes

Retries must remember completed targets and distinguish active
lease holders. Fence state updates, preserve permanent outcomes and
reject missing parked payloads instead of dispatching empty data.

* fix(dispatch): isolate and resume tenant sends

Check tenant sources before filters and never dispatch orphan routes.
Record each target outcome, isolate update chains by forge and group
identity, and keep stale events from overwriting newer messages.

* fix(queue): bound batches and resume deliveries

A queue invocation has shared query and subrequest limits.
Use a small batch and shared send budget, renew fenced leases and
resume only unfinished work without exhausting unstarted messages.

* fix(admin): guard revisions and tenant targets

Read-time authorization alone cannot protect concurrent edits.
Guard revisions, current roles and destination ownership in the
write transaction, and use per-group endpoints for isolated changes.

* fix(webhook): confirm only accepted deliveries

Acknowledging before queue acceptance permanently loses deliveries.
Confirm only accepted work, release failed claims, scope replay
identities and strip tenant-supplied installation authority.

* fix(storage): prune abandoned parked payloads

Oversized payloads can remain after interrupted enqueue attempts.
Prune objects beyond their retry lifetime without deleting payloads
that a current delivery may still need.

* fix(console): expose api conflict statuses

Callers need the HTTP status to distinguish stale edits from other
failures and reload current data after a conflict.

* refactor(console): model group revisions

Carry server revisions explicitly so console writes can reject
changes based on an outdated group snapshot.

* fix(console): send loaded route revisions

Send the revision actually loaded with the route list and adopt the
server response after a save. Reload on conflict instead of silently
overwriting another editor.

* fix(console): guard fragment saves by revision

Fragments share a group version with other configuration writes.
Submit the loaded revision and reload on conflict to preserve
concurrent edits.

* fix(console): save groups with guarded writes

Whole-list saves can overwrite unrelated groups. Use per-group
revision-aware writes, retain rename results and confirm deletion
on the exact group, routes and version shown to the user.

* fix(console): reject edits of stale routes

A conflict reload can leave an older route editor open.
Compare its original value before saving so it cannot overwrite
the newly loaded route.

* fix(console): show overview loading failures

Failed overview requests must not look like an empty dashboard or
a successful refresh. Show the reason while retaining prior data.

* fix(console): explain rejected install bindings

Explain why an installation binding was refused instead of silently
returning to the console with no visible outcome.

* fix(console): allow selecting a group named all

The all sentinel collides with a valid group id.
Use a value that cannot be a group id so that group stays selectable.

* fix(console): discard stale log detail responses

An older detail request can finish after a newer selection or close.
Discard it so the dialog never shows the previously clicked record.

* fix(console): retain legacy group owners

Groups predating member roles store owners in adminIds.
Load those as owners so adding a member never drops all old owners.

* fix(console): lock member edits during saves

Edits made during a save are overwritten by its result.
Disable member controls and guard handlers until the write settles.

* fix(console): show invite loading failures

Failed invite loads must not look like no pending invites.
Show the error and discard responses from a previously shown group.

* fix(console): retain legacy discord log targets

A stored log target without platform means Discord.
Keep that meaning in the editor instead of clearing the target.

* fix(console): retain undisplayed forge sources

An editor that only displays some providers must preserve the others.
Keep stored custom sources when saving unrelated group settings.

* fix(console): lock the group editor while saving

Prevent edits while a group save is in flight and make super-only
installation settings readonly instead of silently ignoring edits.

* fix(console): lock route edits and show failures

Lock the route editor while saving and show fragment load failures
with a retry action rather than presenting a false empty list.

* fix(config): expose the legal contact override

Declare the public runtime key so the documented environment
override reaches legal pages in deployed Nuxt builds.

* build(deploy): migrate d1 before deployment

Apply pending D1 migrations before deploying code that requires
them. Stop deployment when migration fails instead of running
against an incomplete schema.

* docs(worker): align guides with runtime behavior

Synchronize both language guides, API references and examples with
the implemented authorization, storage and delivery behavior.
Document migration requirements and accepted operational limits.

* fix(storage): repair orphaned KV routes

Repair orphaned KV routes before serving KV-only configuration.

* fix(security): fully escape forge HTML

---------

Co-authored-by: RhenCloud <i@rhen.cloud>
2026-10-24 16:19:06 +08:00

21 KiB
Raw Blame History

存储布局

KV 存储布局

KV 只保留缓存数据和短期/临时状态。高频写入(webhook 去重、投递状态、消息追踪)存放在 D1,仅在 D1 不可用或未迁移时才回退到 KV(见存储决策)。

键模式 值 TTL
config:routes / config:groups 旧版(D1 之前)配置,仅在 D1 表为空时读取一次用于播种 永久
config:d1-seeded 标记 D1 已播种(已清空的 D1 不会被再次播种) 永久
session:{id} 管理员会话 { userId, login, loginTrusted } 1 天
token:{userId} { userId, accessToken, expiresAt, live? }(Token 可使用 10 小时) 9 小时
token-reverse:{sha256} 用于按 Token 反查的用户 id 9 小时
state:{hex} { redirectTo, expiresAt, link? }(link:机器人 /gh login 要绑定的聊天账号) 600 秒
link:{nonce} 等待 GitHub 用户确认的机器人账号绑定 600 秒
invite:{token} { groupId, role, expiresAt, createdBy, note?, groupInstance, issuer } 7 天
invite:group:{id} 每组的 Token 索引(保证邀请列表一致性) 7 天
invite:by-group:{id}:{token} / invite:by-instance:{instance}:{token} 每个邀请的列表标记(按分组 id 与签发分组的 instance),并发创建或改名中断时邀请也不会从列表中消失 邀请有效期
login-owner:{login} 第一个使用该登录名的 GitHub 账号 id;之后用同名登录的其他账号只按 user id 匹配 永久
delivery:{provider}:{scope}:{id} Webhook 投递去重回退(D1 dedup_keys 为主);全局入口的 scope 为 @global,分组入口为 @i:{instance}(接收分组不可变的 instance) 2~10 分钟
delivery-state:{provider}:{scope}:{id} 队列投递状态回退(D1 delivery_state 为主) 7 天
queue:payload:{provider}:{scope}:{id}:{uuid} 未绑定 R2 时暂存供队列消费者读取的超大 webhook 负载(每次入队一个键) 4 天
tenant:{groupId} / tenant:i:{instance} 旧版分组 webhook secret,仅在密钥尚未存入 D1 分组行时作为回退读取;轮换或删除时清除 永久
group:renamed:{oldId} 仅 KV 部署:分组改名后的新 id(有 D1 时该别名是改名事务中写入的 dedup_keys 行) 4 天
msg:{chain}:{platform}:{target} 消息追踪回退(D1 message_tracking 为主) 30 天
msg:lock:… D1 不可用时串行化同一消息链更新的锁(有 D1 时是 dedup_keys 行) 240 秒
storage:kv-fallback 标记曾在 D1 表缺失时把状态写入 KV;有效期内 D1 未命中时也会查 KV 35 天
cmd:guilds:{appId} 每个 Discord 应用一条:各服务器已注册命令的哈希与同步游标 永久
cmd:registered:global:{appId} 最近一次全局注册的命令哈希(变化时重新注册,并每天一次) 1 天
config:discord-app-id:{tokenHash} 按 bot token 解析出的 Discord 应用 id 永久
discord-modal:{id} 评论模态框上下文,仅在放不进 100 字符的 custom_id 时写入 1 天
feishu:token:{appId} 每个飞书应用的 tenant_access_token(内存中也会保留) Token 有效期 − 6 分钟
telegram:webhook:{botId} 最近一次 webhook 注册内容的哈希(相同时跳过 setWebhook) 1 天
i18n:{lang} 叠加在英文之上的翻译覆盖 永久
config:identity-migrated 仅 KV 部署:分组身份已由本服务器管理(一次性身份迁移已执行) 永久

D1 存储布局

D1 数据库(DB 绑定,数据库 webhooker)保存配置、投递日志和高频临时状态的权威数据源:

表 用途
d1_groups 分组(权威配置,从旧版 KV config:groups 播种)。每行还保存不会发给控制台的服务端字段:不可变的 instance、legacyId 以及分组 webhook 密钥
d1_routes 每组分组的路由(权威配置,从旧版 KV config:routes 播种)
send_logs 每次分发尝试一行(路由 id、事件、目标、ok/error、耗时、错误码、详情)
audit_logs 每次管理员操作一行(登录/登出、分组/路由/成员/邀请变更)
dedup_keys 原子的一次性占用与记录(INSERT ... ON CONFLICT UPSERT,键 + 过期时间):webhook 投递去重及其确认(accepted:*)、自定义 webhook nonce(nonce:{secret 作用域}:{nonce})、已使用与已撤销的邀请(invite-used:*、invite-revoked:*)、机器人命令/事件去重(bot:{platform}:…)、已撤销的会话与 token 及每个用户当前有效的 token(session-revoked:*、token-revoked:*、token-live:*)、已使用的账号绑定确认(link-used:*)、已退役的分组 id(group-retired:*)与改名别名(group-alias:*)、安装自动配置记录(installation-provisioned:*、installation-bound:*)、排队投递中已完成或已永久失败的路由 × 目标(delivery-target:*、delivery-failed:*)、已发送的分组日志摘要、消息链锁(msg:lock:*)以及一次性初始化标记
delivery_state 队列投递状态(pending/processing/delivered/retrying/failed/dead);领取投递是一次条件 UPSERT(租约),两个消费者不会同时分发同一投递
message_tracking 原地更新用消息追踪(event_id + target_id → 消息 id 及其所显示的版本);消息链按分组 instance 建键,之后以同一 id 新建的分组不会编辑这些消息
d1_fragments 每个分组的命名过滤器片段
feishu_links 映射 feishu_user_id → github_user_id,供 /gh 飞书命令使用
discord_links 映射 discord_user_id → github_user_id,供 /gh Discord 命令使用
telegram_links 映射 telegram_user_id → github_user_id,供 /gh Telegram 命令使用

audit_logs 由定时任务在 AUDIT_RETENTION_DAYS(默认 90)后自动清理。storage-prune 任务会清理过期的 dedup_keys、超过 7 天的 delivery_state 行、30 天未更新的 message_tracking 行(每次编辑都会刷新)以及遗留的 R2 负载。send_logs 不会自动清理。日志行字段说明见日志。

迁移 0013_prune_indexes.sql 增加了这些读取依赖的索引:清理用的 delivery_state(updated_at)、统计页用的两个只含失败 send_logs 行的部分索引,以及替换单列分组索引的 audit_logs(group_id, ts)。索引列每次写入也会多计一行 D1 写入,因此只添加读取节省远大于写入成本的索引。

R2 存储布局

R2(PAYLOAD 绑定,bucket webhooker-payloads)存储对队列消息或 KV 来说过大的 webhook 负载:

对象模式 用途 保留期
webhooks/YYYY/MM/DD/<uuid>.json 暂存供队列消费者读取的超大负载 分发后删除;遗留对象 5 天后清理

队列消息超过 120,000 字节(UTF-8)时暂存负载。当缺少 PAYLOAD 绑定时,超大负载回退到 KV 键 queue:payload:{provider}:{scope}:{id}:{uuid}(4 天 TTL,与队列消息保留期一致)。发送队列消息失败时暂存副本会保留(调用失败时队列也可能已经收下消息);未被使用的副本会过期或被清理。消费者需要时暂存负载已丢失,绝不会以空事件分发:投递保持 retrying,并写入一条 PAYLOAD_MISSING 发送日志,提示从源头重新投递该 webhook。

存储决策

  • D1 是配置与投递元数据的权威数据源;KV 只保存缓存和短期状态。
  • canUseD1 探测(server/lib/storage/d1.ts,检查 prepare + batch)为每个 D1 存储做门槛判定。高频存储(去重及其他一次性占用、投递状态、消息追踪)在其 D1 表尚不存在时(数据库未迁移)回退到 KV。其他 D1 故障会直接抛出——请求返回错误,由发送方或队列重试——因为按调用切换存储会让同一占用分散在 D1 与 KV 两处。
  • D1 配置读取失败时,请求返回错误,由发送方或队列重试;配置没有 KV 副本,分发也绝不会把不可读的配置当作空配置。存储中无法读取的路由或分组行只会让所属分组的投递可见地失败并重试——见路由 → 不可读的路由。
  • 配置写入从不回退:存在 D1 绑定时,写入失败会作为错误返回,而不是暂存到 KV。写入按分组进行(单行,或单个分组的路由/片段在一个事务中),分组行使用比较并交换,不同分组的并发编辑互不覆盖。权限检查、读改写路径、webhook 入口与队列消费者的读取绕过所有缓存(撤销的路由或目标立即停止投递);其他读取在每个 isolate 的内存中最多缓存 60 秒。
  • 旧版 KV 配置只在两张 D1 表都为空时复制一次,且从不覆盖已有行;标记位保证被清空的配置不会复活。
  • 这让 Workers 免费版的每事件 KV 写入趋近于零(每日 1000 次写):去重、投递状态和消息追踪改为写 D1 行(D1 免费版每日可写 10 万行)。
  • R2 免费额度(10 GB-月存储、每月 100 万次 A 类操作)可以轻松承载负载暂存,不占用 KV 写配额。