Covers the docs vs execution split, the four SNOWLUMA_MCP_* env vars, the query_action / invoke_action tools, read/write-mode example client configs, and the per-action safety model (default-write classification, the enforced mode gate, unknown-action rejection, envelope passthrough, and the readOnly/destructive annotations).
@snowluma/mcp gains execution alongside the read-only catalog: when SNOWLUMA_MCP_ENDPOINT is set it exposes query_action (read-only actions, readOnlyHint) and, in write mode, invoke_action (any known action, destructiveHint), proxied to a running OneBot instance over HTTP. Without an endpoint it degrades to docs-only. SNOWLUMA_MCP_MODE (docs|read|write) gates tool visibility AND is re-checked in execute(), so an unlisted tool cannot be called; unknown/legacy actions are non-invokable.
Read vs write is classified precisely at the source: defineAction/preset gain readOnly (default false = write, the safe default), flowed through describe() into the generated catalog snapshot. 58 actions are marked read-only by what their run() actually does -- name-traps like get_ai_record (publishes to the group), request_decrypt_key (returns a DB key), and download_file (arbitrary URL + disk write) stay writes.
The HTTP call is a ~20-line fetch bridge rather than @snowluma/sdk: the SDK's published dist is bundler-only ESM (extensionless imports) that native node cannot load, and the bin stays a plain tsc build.
onebot 332 tests pass (readOnly is metadata-only, no behavior change); mcp gains 21 tests for mode gating / read-write routing / envelope passthrough / error mapping. Full workspace typecheck clean; native node smoke verified.
The four NetworkAdapters each carried a byte-identical reload() state machine (sigChanged/wasEnabled -> Reopened/Closed/Opened/Normal). Lift it into IOneBotNetworkAdapter as a template method; per-adapter variation is now expressed via three hooks: bindingSignature(), willEnable() (default enabled!==false; clients add && !!url), and onConfigReplaced() (post-swap side effects).
Also delete four redundant 'override get isActive()' (identical to the base getter), lift the duplicated uin-logger setup into the base ctor, and drop http-post's now-redundant signature_ cache.
NetworkReloadType is a test-observability seam -- the sole production caller (instance.applyConfigDiff) discards it -- so a new base-level test exercises all transitions directly. Behaviour-preserving: +94/-157 in src, full workspace typecheck clean, 332 onebot tests pass.
A new published package that exposes every OneBot action's docs + a ready-to-use
JSON Schema to LLM clients over stdio. Tools: list_actions / get_action /
search_actions / list_categories, plus the whole catalog as a resource.
The catalog is a build-time snapshot generated from @snowluma/onebot's live
action specs (collectActionDocs()) — so it auto-tracks action add/remove and the
published package carries zero runtime dependency on the private onebot package.
action-kit enrichment (the data source):
- Field.toJsonSchema() per coercer (uint→integer/minimum:1, messageId→integer/
not-const-0, int→min/max, string→minLength/maxLength, array→items, enum→enum).
- describe() composes a per-action inputSchema; ActionDoc gains inputSchema +
category. action-docs.ts tags each doc with its domain category.
Wiring: onebot exports ./action-kit + ./action-docs; root build:all includes mcp;
mcp-npm-publish.yml mirrors the sdk/ui publish workflows (triggers on changes
under packages/mcp/** or packages/onebot/**, since the catalog derives from it).
onebot typecheck + 318 tests green; mcp typecheck/build/smoke green; the 142
registered OneBot action names are unchanged.
Live testing showed the voice-to-text text is delivered by an async push
(Event 0x210 subType 61), not the request response — so the previous
poll-by-resend never got it. Captured a real push and decoded its protobuf:
outer { f1, f2:item }; the item echoes the request msgId at field 1 (the
correlation key) and the recognised text at field 8.
- proto-defs: PttTransPush / PttTransPushItem (live-verified).
- protocol: Event0x210SubType.PttTransResult = 61 + decoder → internal
`ptt_trans_result` event { msgId, text } (replaces the temp hex dump).
- core: translatePttToText now sends ONE trigger and returns the inline text
only if present (already-transcribed); empty = pending.
- onebot: ptt-trans-waiter (keyed by selfUin:msgId), event-pipeline subscribes
`ptt_trans_result` to resolve it; fetch_ptt_text registers the waiter,
fires the trigger, and awaits the push (20s timeout).
Decoded one real push as text "12311123。" end to end.
Live testing showed the voice-to-text text is delivered via an async push
Event 0x210 subType 61 (0x3D), not the sync pttTrans.TransC2CPttReq response
(which is a fixed ~137B ack). That push has no static sys_msg_0x210_0x3d
handler — its protobuf is registered dynamically by the ptt-trans subsystem,
so RE through the std::function callback chain is impractical.
Dump the raw subType-61 body (fromUin + len + hex) once so the wire format
can be decoded from ground truth, then implement the real decoder +
request/await correlation. Temporary — to be replaced by the actual decoder.
Sending got full-chain trace via the api-handler chokepoint; receiving had
none. Wrap convertAndDispatch — the single path every inbound message/notice
takes (decode → convert → dispatch) — as the receive-side mirror:
- When trace is live, run the chain inside a request id so the whole receive
(incl. any rkey-fetch packets the conversion triggers) is correlated under
one [req#N]; otherwise skip the wrap entirely (zero overhead).
- trace the raw inbound event on entry (deep render) and the converted
post_type + elapsed on exit (or "dropped").
Reuses the existing trace/ALS/renderParamsVerbose machinery; no new mechanism.
Debugging from user-supplied logs was impossible: a successful action left
one terse, over-summarized debug line and nothing about what it sent QQ or
what QQ returned. Add full-chain tracing that stays off disk so it scales to
hundreds of groups.
common:
- New `trace` level (below debug). Never written to file — it lives only in
an in-memory ring + the WebUI stream, so the high-volume full chain costs no
disk. `emit` early-returns for trace before format/render (and before any
lazy producer) when trace isn't active → ~free at the default level.
- Dedicated O(1) circular ring for trace (default 5000, SNOWLUMA_TRACE_BUFFER),
separate from the normal 1000-entry buffer (also switched off array .shift);
getRecentLogs merges both by id.
- request-context.ts: AsyncLocalStorage request id. The logger auto-stamps
`[req#N]` on every line in a request's async chain — across packages, zero
signature threading — and records it on the entry.
- log-summary.renderParamsVerbose: deep render of params (nested structure,
per-string cap, total budget, token/password/secret redaction).
chokepoints (auto-cover every action without touching 100+ handlers):
- onebot api-handler: wrap each action in a request id (only when trace is
live); trace the full params on entry and status+elapsed on exit; the terse
debug breadcrumb still persists to file, failures still warn.
- core bridge.sendRawPacket: trace on success (cmd, sizes, elapsed); warn on
errorCode!=0 / no response (cmd, code, message) so disk keeps real problems.
- protocol highway: fold the bare console.log lines into the logger (trace).
webui: trace as the lowest filterable level; render the `[req#N]` tag and let
search match it. Server needs no change — LOG_LEVELS, getRecentLogs and the
SSE stream carry it through.
Adds NapCat-compatible voice-to-text. New `fetch_ptt_text` action (aliases
`get_ptt_text` / `get_record_text`) takes a message_id, resolves the cached
`record` fingerprint (uuid + md5 + duration/size/format), drives QQ's native
`pttTrans.TransC2CPttReq` / `pttTrans.TransGroupPttReq`, and returns { text }.
The response carries the text inline once transcription is done; an empty
text means the server is still transcribing (there is no dedicated push
command — confirmed via RE), so the bridge re-sends with a short backoff
until the text lands, mirroring NapCat's fire-then-re-read shape.
Layers:
- proto-defs/ptt-trans: PttTransReq/Resp + per-scene items.
- core ExtrasApi.translatePttToText: encode -> sendRawPacket -> decode -> poll.
- onebot media-actions.fetchPttText + api-handler/instance-context wiring.
- onebot extended: register the action + aliases.
The commands + request/response field tags were reverse-engineered from
QQ-NT wrapper.node (im_core/msg/ptt/ptt_trans_worker.cc). The numeric tags
are solid; the length-delimited fields (uuid/md5) and the response text tag
are RE-derived and want one live capture to confirm before relying on them.
Introduce packages/onebot/src/action-kit.ts — a zero-dependency,
type-inferring declarative layer for OneBot action params. Each action
declares its param contract (f.* coercers + optional cross-field rules)
and a `run` handler that receives typed, already-validated params.
Validation failures now return BAD_REQUEST naming the offending field
instead of falling through to ApiHandler's INTERNAL_ERROR catch.
- 141/142 actions migrated to defineAction/groupAction/groupUserAction
across info/message/friend/group-*/request/extended; only
.handle_quick_operation stays legacy (it needs the ApiHandler instance,
which run() does not receive).
- f.messageId() distinguishes signed/non-zero message ids from positive
group/user ids (f.uint()); int({min:0}).default(1800) fixes duration=0
(unmute), which the old `|| 1800` silently swallowed.
- run(p, ctx, raw) exposes the raw params as an escape hatch for the
irregular tail (alias-key fallbacks, arbitrary objects).
- action-docs.ts renders self-generated API docs from describe().
- ApiHandler dispatch / try-catch / response-wrapping unchanged (narrow seam).
Registered action name set is byte-identical before/after (142, verified
by runtime enumeration). Typecheck clean; onebot suite green (the
pre-existing websocket native-addon test failure is unrelated).
New "关于" mini-tab in the settings page showcasing SnowLuma: logo +
name + version badge (__APP_VERSION__) + tagline, a short blurb, and a
prominent GitHub Star call-to-action linking to SnowLuma/SnowLuma, plus
secondary links (repo / releases / issues).
Static CTA — no GitHub API fetch, so no rate-limit / loading state.
Reuses /logo.png (same asset as sidebar/login) and the existing Card +
motion + accent-color design language; responsive centered layout.
A hardcoded `#sl` status command — deliberately NOT a plugin system.
Matched at the instance dispatch choke point: a pure-text message equal
to `#sl` (exact, trimmed, case-insensitive — no startsWith) replies with
SnowLuma version / platform / uptime in the same conversation.
- Anyone can trigger; the per-UIN `statusCommand.enabled` toggle is the
only gate (default on). Per-conversation cooldown (default 5s,
configurable) prevents flood/风控 from repeated triggers.
- `swallow` (default false / pass-through): when set, a matched `#sl` is
not forwarded to downstream adapters — still cached, logged, replied.
- Config on OneBotConfig.statusCommand (per-UIN): defaulted + clamped in
config.ts, mirrored/normalized in the WebUI; edits hot-reload via the
shared instance context.
- WebUI: a "内置命令" section in the 通用设置 tab (enable / swallow /
cooldown). Node enable toggle extracted into a shared ui/ToggleSwitch.
Pure matcher / cooldown / uptime / text helpers live in
modules/status-command with unit tests; config defaults + clamp covered.
/get_forward_msg returned raw, rkey-less image/media URLs because
getForwardMessage converted forward nodes WITHOUT the URL resolvers the
normal receive path (to-message.ts) uses. Wire imageUrlResolver +
mediaUrlResolver into it so forwarded images/voice/video get signed
download URLs — exactly issue #74 (image url 缺少 rkey).
Also derive the image rkey scene from the appid baked into the URL
(1406 -> private, 1407 -> group) instead of the carrying message's
group/private context, falling back to isGroup only when the URL has no
recognised appid. A c2c image forwarded into a group keeps its original
c2c appid, so the old isGroup-based pick signed it with the group rkey ->
`invalid rkey` on download. Mirrors NapCat's getImageUrlFromParsedUrl;
Lagrange/NapCat both reuse one unified converter so this never drifts.
Fix the self-contradictory `c2c images` rkey test (its fixture used the
group appid 1407) and add cross-tests asserting appid outranks isGroup
both ways, plus the no-appid fallback.
PR #73's login-state probe shipped with issues:
- `fetchPtlogin` returned `Promise<any[]>`, tripping the repo's
`@typescript-eslint/no-explicit-any` (error) and breaking `pnpm lint`.
Modelled the read fields as a `PtloginUin` interface and cast the parsed
JSONP to it.
- `tryPtloginMethod` had a dead `if` whose two branches both returned
'fallback'; collapsed to a single return with a comment.
- The `< 6` QQ-process-count heuristic was a bare magic number (one copy
annotated only with "13579 5个端口全部用光"). Extracted to
`LOGGED_OUT_PROCESS_COUNT_MAX` with an explanation and reused in all three
sites (incl. the error fallback).
Behaviour-preserving: the probing logic (Ptlogin parse, deep-link fallback,
PID port discovery) is unchanged — only typing and the magic constant.
set_group_whole_ban false (unmute) failed with `OIDB error 1007` while
enabling worked (issue #70). OIDB (0x89A, 0) is shared by several commands
(SetSearch / SetAddOption / mute-all) disambiguated by body shape. Unmute
sends muteState.state=0, but proton omits zero scalars (proto3), leaving an
empty `muteState` the server can't classify. Enable sends 0xFFFFFFFF, which
is non-zero and survives — hence the asymmetry.
Add a `pb_optional<N, T>` marker (proto3 explicit presence): a scalar so
marked serialises even at its zero/default value. It is opt-in per field via
a NEW `explicitPresence` flag set ONLY by this marker — never reused from the
TS-`?`-derived `isOptional` (true for nearly every field). The encoder routes
explicit-presence scalars to a dedicated block; every `pb<>` / `pb_repeated<>`
field takes the untouched branches, so existing wire output is byte-identical
(proton's 99 existing tests, incl. byte-level round-trips, stay green).
Mark OidbMuteAllState.state `pb_optional<17, uint_32>` so unmute now puts
field 17 (=0) on the wire. New tests: proton pb_optional encode/round-trip
(incl. that sibling pb<> fields still omit zero), and a byte-level assertion
that mute-all unmute carries the state field.
Note: proton ships as a built Vite plugin (dist, gitignored); rebuild it
(`pnpm --filter @snowluma/proton build`) for the change to reach consumers.
set_group_card always failed with `OIDB error 1007 on 0x8fc_3` (issue #70).
The 0x8FC request protos put the body wrapper at tag 2 and the member card
name at tag 2, but QQ-NT expects body at tag 3 and targetName at tag 8, so
the server found neither and rejected the request.
Cross-checked byte-for-byte against two working references:
- Lagrange.Core OidbSvcTrpcTcp0x8FC: Body=ProtoMember(3), TargetName=8
- NapCat Oidb.0x8FC_2: body=ProtoField(3), targetName=8
Fixes OidbRenameMember (0x8FC_3 set card) and the same latent bug in
OidbSpecialTitle (0x8FC_2 set special title), whose body was also at tag 2.
Send-only protos (empty deserialize); existing tests assert the command
name, not field tags, so they're unaffected.
The homepage was doing double duty — monitoring + injection control. Split it:
- New /processes page ("进程注入", sidebar #2) owns the QQ-process control
surface (load / unload / refresh / probe-login + its dialogs). Overview is
now a read-only dashboard: stats, host metrics, online accounts, plus a
read-only injection-summary tile (online / loadable, links to /processes)
and a first-run CTA that points at /processes while nothing is online.
Route + AppPath + nav item wired; state/polling stay layout-owned.
Responsive + a11y pass:
- Mobile sidebar Sheet gets an sr-only title + description, fixing the Radix
"DialogContent requires a DialogTitle" accessibility error.
- Log rows stack on phones (meta line + full-width message) instead of a
~40px column that broke every few chars; time parses defensively so a
non-ISO value never renders "Invalid Date".
- Process cards stack on phones (info row + actions row) so name / PID / path
stop collapsing to "Q.."; the status badge no longer wraps.
- Config tab strip scrolls horizontally on phones instead of wrapping to
three rows; account strip defaults collapsed ≤lg; header buttons wrap.
- Dialogs keep a 1rem gutter on tiny screens via w-[calc(100vw-2rem)].
- Content widens to 1600px on ≥2xl; account/process grids gain a column at
2xl so >1700px screens aren't mostly empty gutter.
Add a new animated SVG logo at .github/logo.svg and embed it in README.md. Update the README content: refresh badges, simplify and translate the project intro and feature list into Chinese, add concise quickstart steps, adjust links (Releases → 发布版, RoadMap), and replace long/duplicated sections with a shorter overview and contributor badge. Overall this cleans up and modernizes the README and adds a branded logo for the project.
Five WebUI changes:
1. Settings page rebuilt around mini tabs (外观 / 数据刷新 / 账号安全)
instead of three stacked cards; content and behaviour unchanged.
2. "修改访问密码" now opens an independent ChangePasswordDialog modal
rather than swapping the whole page for the forced-change screen. The
form is extracted to a shared ChangePasswordForm used by both the modal
and the forced full-screen ChangePasswordPage — behaviour identical.
3. Node config: creating / editing / deleting / enable-toggling an
HTTP/WS endpoint now persists immediately — clicking 保存 in the editor
dialog IS the save. Previously it only marked the config dirty until you
also hit the top-right 保存, which silently cost many users their token
edits. `save()` takes an explicit config so it persists the freshly
computed value, not the stale closed-over state. The general-settings
tab keeps its explicit top-right save.
4. Forced first-time change-password no longer renders an old-password
field: it carries the password the user just logged in with. Browsers
(Edge/Chrome) were autofilling that field on upgrade, so users thought
the old password was pre-filled correctly and could never save. A
returning already-authed session (no fresh login) still shows the field.
5. Added default error + 404 screens (status-screens.tsx), wired into the
TanStack router via defaultErrorComponent / defaultNotFoundComponent.
Addresses MliKiowa's review on PR #68 ("单独排除 sqlite 就行了"). That PR
fixed the `DatabaseSync is not a constructor` crash by switching the core
bundle to SSR mode (build.ssr + ssr.noExternal + target node22 + a
catch-all isNodeBuiltin external fn). The root cause was narrower: on
Node 22.x `node:sqlite` is experimental and absent from `builtinModules`,
so the `nodeModules` sweep didn't externalize it and Vite swapped in a
browser shell. The SSR switch also self-inflicted the @hono/node-server
MODULE_NOT_FOUND it then had to paper over with noExternal — the original
config already bundled all non-builtin deps.
Revert vite.config.ts to the pre-PR state and just add 'node:sqlite'
(+ bare 'sqlite') to the external list. Verified: `SNOWLUMA_TARGET=linux-x64
vite build` succeeds and dist/index.mjs now carries a real
`import { DatabaseSync } from "node:sqlite"` rather than a shimmed shell.
get_image and get_msg returned the URL baked into the message store when
the message first arrived. That URL's rkey expires, so a client fetching
it later (or re-querying after the original push) got
`{"retcode":-5503010,"retmsg":"invalid rkey"}`.
Re-sign at lookup time from the raw unsigned `imageUrl` the media store
already keeps alongside the signed `url` (plus `isGroup`): getImageInfo
now takes the image rkey resolver and mints a current rkey, and get_msg
walks its segment array refreshing each image URL through it. Both are
best-effort — a resolver miss/throw or a non-array message body falls
back to the stored URL untouched. Mirrors NapCat re-signing on every
getImageUrl call instead of handing back a stale baked URL.
The resolver is reused from converterCtx (no new rkeyCache wiring); a
cache-hit re-sign is a string concat, only an expired rkey awaits the
already-deduped/cooled-down/timeout-bounded refresh.
Image/video/record download URLs are signed at receive-time with an rkey
(OIDB 0x9067_202, TTL-bounded). RKeyCache resolved synchronously: on a
cache miss or inside the expiry skew it returned the URL *without* an
rkey and only kicked off a fire-and-forget refresh (plus a 30s cooldown).
So during the expire→refresh window — and whenever the OIDB hiccuped —
SnowLuma emitted unsigned URLs, which the CDN rejects with
`{"retcode":-5503010,"retmsg":"invalid rkey"}`. Intermittent by nature.
Mirror NapCat's getImageUrl (`if (isExpired()) await refreshRkey()`):
make resolveImageUrl/resolveMediaUrl async and block on one refresh when
the cache can't serve a still-valid rkey. The refresh coalesces
concurrent callers onto a single OIDB round-trip, backs off for 30s after
a failure, and is bounded by a 5s timeout so a stuck packet send never
wedges message conversion (it degrades to whatever the cache holds).
Widen ImageUrlResolver to `string | Promise<string>` and await it in
to-segment so the image path can resolve asynchronously like media.
loadBinarySource issued a header-less fetch(source) for remote media.
Anti-hotlink / anti-bot front-ends reset such requests, surfacing as
undici `TypeError: fetch failed` (not a clean 403) — so quoting a
command and sending an image/GIF whose source is a guarded URL bounced
with send_group_msg retcode 1200 "fetch failed".
Mirror NapCat's download path: send a browser User-Agent on every
request and retry once with a Referer pointing at the resource when the
first attempt is refused (network error or 403/4xx). Size-limit
rejections are tagged noRetry so an oversized body isn't re-downloaded.
Also route set_group_notice's image fetch through loadBinarySource so it
inherits the same hardening (plus http/file/base64 handling).
Remove the in-tree @snowluma/sqlite wrapper and vendored better-sqlite3 prebuilts; migrate consumers to node:sqlite's synchronous API. Replaced imports and types (Database/Statement -> DatabaseSync/StatementSync), updated PRAGMA calls to use exec, and adjusted code in onebot (media/message/reaction stores) and protocol (identity service). Vite config no longer bundles sqlite native binaries and references to the sqlite workspace were removed. Also delete the packages/sqlite package and its native prebuild artifacts, and add a couple of unrelated AST improvements in proton: use a module-level shared wrapper cache and add an inheritance cache to avoid repeated scans. Minor import reordering/cleanup applied where needed.
Reorder and normalize imports, add missing type imports, and fix whitespace/formatting across multiple files. Notable updates: eslint.config.ts import ordering; core bridge APIs (group-album, misc, web) now import Json types; tests/highway pipeline import ordering and added type imports; protocol changes (element-builder, highway client/pipeline) reorganized imports and added proto type imports; many oidb service modules moved protobuf encode/decode imports for consistency; web group modules (album, essence, honor, notice, request-util) reordered imports, adjusted typings/indentation, and small whitespace fixes. These edits are stylistic and typing improvements and should not change runtime behavior.
Enable @typescript-eslint/no-explicit-any as an error in the ESLint config to tighten typing across the codebase, and add an override to disable this rule for test files. Also apply non-functional cleanup in packages/proton/src/runtime.ts: reorder/imports and adjust formatting/indentation and string alignment for readability — no behavioral changes.
Remove an unused Bridge import, the asBridge helper and a stale comment from group-album, and relocate the GroupAlbumMediaResult and GroupAlbumCommentResult interfaces to the end of the file. In the sqlite package, remove the inline eslint-disable-next-line for @typescript-eslint/no-explicit-any near the BetterSqlite3 constructor. These edits clean up dead/stale code and re-enable the lint rule; there are no behavioral changes.
Reorder and tidy imports, re-export types, and apply minor formatting fixes across the bridge package. Changes include:
- Reordered imports and grouped related imports in hook-manager, hook-session, pipe-watcher, qq-hook-client, qq-port-probe and others.
- Moved several type re-exports (HookProcessBaseInfo, QqPortLoginInfo, HookProcessInfo/Status) in hook-manager and consolidated exports in package entry index.ts.
- Minor whitespace/style tweaks: constructor spacing, for-loop formatting, empty catch handlers, and small refactors that do not change runtime behavior.
These are non-functional housekeeping changes to improve consistency and readability.
Move and standardize API implementations under bridge/apis, slim down the Bridge class (remove inline action implementations, caches and large comment blocks) and relocate related types/exports. Clean up and reorder imports across many bridge/api modules, normalize constructors, add logger imports where needed, and adjust WebUI routing/import ordering. These changes centralize API logic, reduce Bridge surface area, and tidy module boundaries and imports.
Remove extraneous comments, redundant explanatory notes and extra blank lines across onebot packages; reorder and normalize import lists in group-admin and group-album; remove unused comment blocks in event-converter. These are non-functional cleanup changes in packages/onebot/src/actions/extended.ts, packages/onebot/src/actions/group-admin.ts, packages/onebot/src/actions/group-album.ts and packages/onebot/src/event-converter/index.ts to improve readability and maintain consistent style.
Add JsonObject import and use it for handleSendPacket param typing; replace deprecated hex.substr with hex.slice; convert and consolidate many English docstrings/comments into Chinese for clarity; remove some JSDoc noise and minor formatting tweaks; improve explanatory comments around streaming download, base64 size checks, safe path resolution and async file writes. These are primarily documentation and small type/cleanup changes with no functional behavior changes.
Split the QQ NT client lifecycle / hook injection layer out of
@snowluma/core into its own self-contained package. The new bridge
package owns:
- native hook addon loader (injector)
- HookManager + per-PID HookSession orchestration
- PipeWatcher for QQ process discovery
- QqHookClient pipe protocol
- qq-port-probe for early UIN sniffing
The hook layer used to type its main collaborator as
`BridgeManager` directly. To keep the new package free of any
dependency on @snowluma/core, the cross-package coupling is now
expressed via a 3-method `BridgeManagerSink` interface that lives
in the bridge package. BridgeManager (in core) satisfies the
interface structurally with no source changes; the top-level
app/webui still inject it the same way.
Pure file relocation otherwise: git mv preserves history at 100%
on every file, the only logic edit is the interface introduction
in hook-manager.ts and an updated `nativeSearchDirs()` offset
(one fewer `..` since hook now lives at packages/bridge/src
instead of packages/core/src/hook).
The package previously named "bridge" houses the protocol-byte layer
(OIDB cmds, msg-push decoders, highway uploaders, element-builder,
event bus). The name "bridge" suggested transport/lifecycle, which it
isn't — that role lives in core/hook and core/bridge/manager. Rename
to match what's actually inside.
This is a pure rename: zero logic changes, all 73 import sites updated
mechanically and TS-verified. core/onebot dependencies key updated.
ESLint override path adjusted for the moved oidb-services directory.
Rename detection threshold: git log --find-renames=20% recommended
when following history across this commit — content-replacement
(@snowluma/bridge → @snowluma/protocol) dropped per-file similarity
below git's default 50% on import-heavy files.
User report:
SnowLuma's group_invite event renders to OneBot with empty
`comment`, so the bot's approval-prompt template shows nickname
+ uin + flag but loses the requester's verify message.
NapCat's same flow shows the verify text:
【进群申请】批准/驳回:
昵称:KitaIkuyo
QQ:1957003260
flag:1779543612823906
等级:7
你们好 ← only NapCat shows this line
Root cause:
The OIDB push (GroupRequestJoinNotice subType=84) only carries
`(groupUin, targetUid)` and a few opaque numeric fields — no
string for the verify text. NapCat doesn't use the push directly;
they poll `nodeIKernelGroupService.getGroupNotifies` (= our OIDB
0x10C0 pending-request queue at `FetchGroupRequests`) which
returns each pending row with `postscript` (the verify text) and
`seq` (server sequence id). Reference:
dev/NapCatQQ/packages/napcat-onebot/index.ts:483-498
(`flag = notify.seq`, `comment = notify.postscript`).
Our pipeline didn't enrich. The decoder set `message: ''` and
emitted, so OneBot's `comment` was empty.
Fix:
- Extend `PacketPipelineDeps` with a `resolveGroupJoinRequest(
groupId, uid, subType)` hook. Bridge.ts wires it to
`fetchGroupRequests` and matches on `(groupId, uid)` —
targetUid for `subType='add'`, invitorUid for
`subType='invite'`.
- `dispatchAfterStrangerResolve` now fires the profile + request
lookups in PARALLEL via `Promise.allSettled` so a flake on one
path doesn't kill the other. The request lookup's `comment`
lands on `event.message`; OneBot converts that to `comment`.
Decisions:
- Kept the existing `add:<groupId>:<uid>` flag format (consumer
bots already parse it). Switching to NapCat's sequence-form
flag would break callers and is a separate breaking change.
- Both lookups fail-open: if either rejects, the event still
emits with whatever info we did get, so the bot at least sees
the request and can render the bare nickname.
Regression tests:
- `populates event.message from the pending-request queue` — pins
the user-reported "你们好" case end-to-end through the pipeline.
- `runs lookups in parallel` — profile resolves, request fails;
event still ships with the resolved uin and an empty message.
- `invite subtype matches the request on invitorUid` — explicit
contract for the invite-forwarded-to-admin case where the
pending row's REQUESTER is `invitorUid`, not `targetUid`.
345 bridge tests pass (was 342 + 3 new); typecheck + lint clean.
User report: after [a4909ec], log STILL shows
群邀请 950929451 -> [bot group(950929451)]
get_stranger_info params=user_id=950929451 → 未知昵称
But NapCat and a sibling SnowLuma bot in the same group BOTH decoded
the same push as `QQ:1957003260 nickname:KitaIkuyo`. So the wire is
correct — our bot is just looking it up wrong.
Root cause: **legacy cache pollution**. The pre-fix decoder fallback
stored `<requester_uid> → <groupUin>` in the identity DB whenever
`resolveUidToUin` missed (because the fallback `ctx.fromUin` on a
group-scoped push IS the group's own uin). After upgrading to
[a4909ec], `decodeGroupJoinRequest` calls `resolveUidToUin` which
hits the polluted in-memory map (re-populated from SQLite at boot)
and returns the groupUin instead of 0. The new `needsStrangerResolve`
guard only fired on `fromUin <= 0` — so the polluted value flowed
straight through and re-emitted the bug.
Two-part heal:
1. `needsStrangerResolve` now ALSO fires when `fromUin === groupId`
— the pollution signature. A real requester's uin can never
equal the group's own uin, so this is a safe override that
forces the async UID-form FetchUserProfile lookup, which then
`rememberUidUin`s the correct mapping on success (eviction is
already wired in identity-service.ts:468-471).
2. `rememberEventIdentity` for group_invite now passes `uin: 0`
when `event.fromUin === event.groupId`, so future writes don't
re-pollute the cache (`rememberUidUin` short-circuits on
normalized uin ≤ 0).
Cache self-heals on first event after upgrade — no manual DB nuke
needed.
Regression test pins the legacy-pollution path: feeds an event
where `fromUin === groupId` (matching the user's log), verifies the
resolver fires and patches fromUin to the correct value.