main
35
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
|
7172ab376d |
Feature: add build_client MCP tool and improve build dialog linkage
Add a service-local MCP tool, build_client, that patches a client template with this master's connect settings and writes the binary on the server. It covers ghost/TestRun/ServerDll/TinyRun plus Linux, macOS and Android, reusing the same byte path as CBuildDlg::OnBnClickedOk (CONNECT_ADDRESS overwrite, build-date check, embedded TLS anchor before UPX/padding). ShellCode AES and payload impersonation stay dialog-only. The tool is gated by a new McpBuild switch (default 0) and McpReadonly=0. It refuses to overwrite an existing file unless overwrite=true, protects the server binaries/settings.ini/upx.exe, validates names strictly and uses a unique UPX temp file so it cannot race the dialog. Results report the path, size, target, TLS-anchor status and follow-up tips. In the build dialog, Linux/macOS/Android targets now also grey out the install-location combo/label and run mode, save and restore the Windows compress choice, hide payload/download rows, grey the random-name and run-as-admin menu items, and fix the TLS-anchor label for posix targets. OnBnClickedOk is untouched; generated output is unchanged for every target. New UI strings were added to the en_US, zh_TW and de_DE language packs. McpBuild is off by default and changes require a server restart. Verified on a Debug x64 instance: MCP generation for Linux is patched correctly, Android and ghost anchor/error paths behave, and dialog-driven Linux/TestRun/SC_AES generation still succeeds. mcp/tools.json regenerated (55 tools); check_platform_table.py passes. Co-Authored-By: deepseek-flash |
||
|
|
06b93f7bdb |
Feat(client): complete the macOS feature set (phase 0 and 1)
Phase 0 makes implementations that already existed reachable. macOS now declares TERMINAL, SCREEN_CLIPBOARD, PROCESS and C2C_RECV_ACK; PROCESS is a new bit and Linux declares it too. SystemManager was dead code -- its constructor took a second argument and it did not derive IOCPManager, so the sub-connection skeleton could never instantiate it -- and now matches the createHandler/OnReceive contract, so COMMAND_SYSTEM really spawns a sub-connection (proc_listpids, SIGKILL). DataProcess always gets a NULL user, so main.mm keeps a g_pMainClient handle and the V2 C2C receiver finally answers TOKEN_C2C_RECV_RESULT. The server asks for SupportsTerminal() on MAC for exec_command / terminal_* and for SupportsProcess() for kill_process, so an old build gets "client is too old" instead of a command nobody answers; kToolPlatforms and mcp/tools.json follow. http_probe stays win/lnx: the terminal works on macOS, but the probe uses GNU-only base64 -w. Phase 1 is POSIX parity. ScreenPreview captures through CGDisplayCreateImage, scales and ships BGRX for the server to encode, and the capability is only declared when preflight or a real capture says the display is readable. ActivityHistory samples the frontmost layer-0 window once a second and prints the same records as Windows and Linux. COMMAND_SCREEN_KEYFRAME_REQ is now consumed by Linux (x264 IDR) and macOS (VideoToolbox force keyframe): the receive thread raises a flag, the capture thread -- which owns the encoder -- consumes it and puts it back if the frame never went out. The macOS quality level persists in NSUserDefaults because main.mm rewrites config.conf. Linux client and Debug x64 server rebuild green; check_platform_table PASS. Not built or run on a Mac. Co-Authored-By: deepseek-flash |
||
|
|
d49ebe521d |
Docs(mcp): correct claims in the server instructions agents rely on
The initialize.instructions text is the only guide a model gets (it cannot read docs/), and several statements disagreed with the code: - get_keylog was described as always visible; it is gated by McpKeylog and hidden when the switch is off. - Dangerous machine actions were not mentioned: freeze/thaw and stop/restart stay in tools/list but need McpMachineDangerous and return -32006 at call time. - -32006 was described as policy/admin-only; MFC also returns it when an old client lacks CLIENT_CAP_V2. The Go server returns -32005 for that case, so its own instructions now say so. - -32001 advice assumed every failure is a timeout; it also covers cancellations, http_probe parse failures and unconfirmed writes. - "Unsupported combos always return -32005 immediately" ignored paths that surface as -32001 once a sub-connection is involved. - Linux matrix listed all host_* as unsupported, but host_set_remark and host_set_watch_level only touch server-local data. - local_dir / local_path are paths on the controller, not the user's box. Also align the error table and platform distribution in the docs, and note that hidden fallback aliases remain callable. Debug build green; go test, check_platform_table and gen_go_tools --check pass. Co-Authored-By: deepseek-flash |
||
|
|
292c3191dc |
Fix(mcp): make declared schemas match what the handlers return
Several tool contracts disagreed with the runtime shape, which forces an
agent to guess instead of following outputSchema:
- get_server_status returned {"status": {...}} while outputSchema has the
fields at top level; return the status object directly.
- get_screenshot description claimed structuredContent.image carries the
base64 data; it only carries mimeType/width/height/bytes, the bytes are
in the content image block.
- host_control_service accepted service_name silently, but not in schema;
document the alias and add anyOf [{required:[name]},{required:[service_name]}]
so a call missing both fails at validation, not at execution.
- kill_network_connection dry_run invented dryRun/connection fields absent
from ActionOutputSchema; use the shared BuildDryRunResult shape.
- search_file outputSchema left files.items untyped; declare
path/name/modified/dir/size and the paging fields it already emits.
- chat_* id hints still named the removed list_online_hosts/search_hosts.
tools.json hunks keep the live-dump key order (alphabetical inside schemas),
tools_minimal.json is regenerated, and mcp_smoke.ps1 now reads the flat
structuredContent. Debug build green; check_platform_table, gen_go_tools
--check and go test ./... pass.
Co-Authored-By: deepseek-flash
|
||
|
|
d4f10db1f2 |
Feature(Go): verify_tools --subset and MCP deployment docs
Add a third mode to verify_tools.py that validates the Go server end to end: it fetches /mcp tools/list and requires the name set to equal the nine implemented tools exactly (seven in readonly mode), compares description / inputSchema / outputSchema field by field against mcp/tools.json, and also checks the published order. Unlike --live, extra or missing tools are failures, not notes. The field-comparison loop is factored into report_field_diffs so --offline / --live and --subset share one comparison, diff printer and FIELDS tuple. Document and deploy the endpoint: the systemd env example gains YAMA_MCP_TOKEN / READONLY / TERM_IDLE / MAX_CONCURRENCY / READ_MAX, the design doc moves to implemented and its section 10 now matches the code (embedded tools_minimal.json, the three packet dispatch groups). The tools README explains gen_go_tools.py --check and --subset. Verified against a freshly built debug Go server on 16543/18090: read-write mode passes with 9 tools, YAMA_MCP_READONLY=1 passes with 7, and both mismatch directions (extra tools, missing tools) fail with exit 1. gen_go_tools.py --check stays clean; no Go code changed. Co-Authored-By: deepseek-flash |
||
|
|
3f1957a9c9 |
Feature(Go): MCP endpoint with bearer auth and tools/list
Add a stateless POST /mcp endpoint to the Go server, mounted by the web server and built from an environment-driven Config in cmd. The endpoint gates on YAMA_MCP_TOKEN: without it the route is not registered and /mcp 404s. Requests use a constant-time bearer compare, and transport failures stay transport failures (401 with a WWW-Authenticate challenge, 405 for anything but POST, 413 above 8 MiB, 204 for notifications). Everything after a parsed JSON-RPC request is HTTP 200, including -32600 for batch arrays and -32601 for unknown methods or tools. tools/list returns the nine minimal tools verbatim from an embedded copy of the canonical schemas. tools/mcp_tools/gen_go_tools.py extracts them from mcp/tools.json, preserving order and only the four published fields, and its --check mode detects drift. Readonly mode hides and refuses terminal_exec/terminal_close (-32006); tools that are listed but not implemented yet answer -32000 until their milestone lands. Also sync design §3.2 and the M2/M7/M8 plan text to tools.json, which is the sole schema contract: files not entries, line-based read_file, no since, camelCase capabilities, max_width [64,1920] defaulting to 1024, and outputSchema included in the M8 subset comparison. Co-Authored-By: deepseek-flash |
||
|
|
5841f27df9 |
Feature(mcp): unify window and service controls behind one name per state
operate_window / host_control_window and control_service / host_control_service were two implementations of the same intent. tools/list now exposes exactly one of each pair, chosen by McpMachine: the host_control_* names when host management is enabled, the main-connection names when it is not. The advertised tool accepts the union of actions. host_control_window routes close/show to the main-connection implementation, which stays the only path with close/show semantics and requires McpManage; the machine path keeps its post-action verification for restore/hide/maximize/minimize/activate. host_control_service routes set_auto/set_manual the same way and rejects dry_run for them instead of pretending to run one. Both old names remain callable aliases, and service_name is accepted as an alias for the machine tool's name parameter. The manifest generator treats the two main-connection tools as !machine entries, so the two-capture union still covers both states. tools/list is 49 entries when everything is on (was 51) and 47 when host management is off; the response shrinks from 193,008 B to 168,634 B (-12.6%), short of the 15% estimate because state-conditional tools must stay in the manifest. The plan document records that measurement and the reason. Verified with a dedicated two-state test: visibility matrix, close/show and set_auto/set_manual routing through dry_run and validation paths, alias dispatch, and -32006 for host_* names when host management is off. Step 8a/8b regressions and verify_tools --live still pass. Co-Authored-By: deepseek-flash |
||
|
|
c73db530f4 |
Feature(mcp): hide the degraded list_* tools when host management is on
host_inspect now covers processes, windows and services, so with McpMachine=1 the older list_processes / list_windows / list_services entries only competed with it: same intent, two names, two long descriptions to read before choosing. They no longer appear in tools/list while host management is enabled, but their names still dispatch, so callers that cached them keep working. They must stay visible when McpMachine=0: that is the default state and the only way such a server can list processes, windows or services, so the entries sit behind !IsMachineEnabled() rather than being removed. list_network_connections is deliberately not folded in: it reads the connection-manager view (IPv6 rows, killable flag, numeric state) which host_inspect's IPv4-only network aspect does not provide. tools.json is now generated from two captures, all-switches-on plus McpMachine=0, because neither state contains the whole surface; the off-state entries are marked gate "!machine" and verify_tools.py skips them when comparing against an all-on server. tools/list went from 54 to 51 entries and from 183,336 B to 171,929 B (-6.2%). Co-Authored-By: deepseek-flash |
||
|
|
59f0f16f74 |
Feature(mcp): merge the ten host_list_* tools into host_inspect
The ten read-only host-management tools (host_list_processes, windows, network_state, software, browser_history, favorites, win32_services, driver_services, scheduled_tasks, host_get_hosts_file) each carried a name, description and schema for what is one connection to the host, one tab id per aspect and one paged list shape. Ten entries made tools/list hard to scan and took a large share of the description budget. They are now one tool, host_inspect(aspects[], ...), where aspects is a required, deduplicated array of the ten former names. A single aspect returns exactly what the old tool returned (same builder, same code path), so output is byte-identical; several aspects are fetched over separate managed sessions and merged into results.<aspect>, with failures isolated in errors.<aspect> and isError only when every aspect failed. Filters (pid/name/arch/title/url/publisher/folder/status/ start_type/process/protocol/ports/state/max_bytes) are still applied server-side before paging and are ignored by aspects they don't apply to. The old names stay as dispatch aliases so existing callers keep working but no longer appear in tools/list; the platform table, batch read-only list, tool hints, instructions, generator tables and the captured mcp/tools.json move to the new name (63 tools -> 54). The description also notes that win32_services / driver_services need an elevated client: a non-elevated one cannot open SCM and never answers, so the call times out after ~12 s. Co-Authored-By: deepseek-flash |
||
|
|
88269df548 |
Feature(mcp): merge the three host-list tools into list_hosts
list_online_hosts, search_hosts and list_all_hosts differed only in whether offline hosts were included and in which filter fields they accepted, which made the host-list surface confusing and duplicated three handlers over the same CollectAllHosts snapshot. They now share one tool, list_hosts(query?, online?=true, include_offline?=false, group?, os?, client_type?, limit?, offset?). The old names stay as aliases routed to the same handler, so existing callers keep working, but tools/list exposes only list_hosts. The legacy "offline" boolean of list_all_hosts is still honored, and list_all_hosts keeps its online+offline default. With offline history included the list is now ordered online-first: the host library holds thousands of records, so the default page of 100 would otherwise be all stale entries and the online hosts a caller can actually operate on would be buried behind paging. instructions, every id-description string, the platform table, the tools.json generator and the captured mcp/tools.json are updated to the new name (65 tools -> 63). Co-Authored-By: deepseek-flash |
||
|
|
613aa7291a |
Feature(mcp): add http_probe for private-network HTTP checks
Run an HTTP/HTTPS request inside a shell on a managed host and return a structured result: status code, headers, body (UTF-8 or base64), size and elapsed time, so interface health on the host or its subnet can be checked without a full remote-control session. The safety boundary is hard-coded on the server, not in the shell script: only http/https and IP literals (or localhost); only loopback and private ranges (127/8, ::1, 10/8, 172.16/12, 192.168/16) are accepted; redirects are not followed; the body is capped at 64KB; TLS certificates are verified. Domain names are refused on purpose - resolving them on the server would probe the wrong network. Windows runs a single-line PowerShell 2.0-compatible script based on HttpWebRequest. Accept/User-Agent/Referer are mapped to their dedicated properties because .NET rejects them through Headers.Add. Linux uses curl with temp files. Payloads cross the terminal channel as 56-character base64 chunks with a declared byte length. ConPTY re-emits the boundary character of lines wider than 80 columns, so single-line base64 arrives corrupted; the length check converts any remaining loss into an explicit error rather than silent garbage. The tool reuses an existing persistent terminal session. Sessions it creates itself are closed after the command and kept only when the probe timed out while still running. Platform gating: Windows and Linux only. Co-Authored-By: deepseek-flash |
||
|
|
697044d84b |
Feature(mcp): add read_file for byte-accurate remote text reads
read_file fetches a single remote file through the existing V2 download channel into a per-call temp directory, decodes it, and deletes the temp files on every exit path. It returns the file size, detected encoding, EOL style, total line count, the requested slice (negative offset counts from the end), a truncation flag and the whole file's SHA-256. Encoding auto-detection follows BOM -> UTF-8 validity -> GBK fallback, and utf-8/gbk/gb18030/latin1 can be forced. Files above 2MB and directories are rejected from the first chunk, so nothing is written to disk for them. The tool is gated by McpFileTransfer only: it never writes to the remote host, so it stays available in readonly mode. The V2 receive path now serves both download_file and read_file. Co-Authored-By: deepseek-flash |
||
|
|
b63c97c5eb |
Feature(mcp): allow sc qc/enumdepend/sdshow in default command whitelist
exec_command's read-only default whitelist only listed `sc query`, so service configuration inspection was rejected even though `sc qc` is a read. Add `sc qc`, `sc enumdepend` and `sc sdshow`; `sc start/stop/config/create/delete` stay rejected because matching is prefix based, and a comment says so. verify_tools.py --live now prints the tools/list response size plus the description / inputSchema / outputSchema breakdown, so shrinking tools/list can be measured against a running server. Verified on an isolated Debug instance (port 6555, local test client): sc qc/enumdepend/sdshow/query succeed; powershell, icacls, net user, net localgroup, sc start and `&` chaining still return -32007/-32012; tools.json still matches the live tools/list (63 tools, PASS). Co-Authored-By: deepseek-flash |
||
|
|
a801a2999a |
Feature: Support input blocking and privacy overlay on Android hosts
remote_block_input rejected anything but Windows because the Android client had no COMMAND_SCREEN_BLOCK_INPUT handler. This adds the overlay on the client and the server gating, and wires up COMMAND_SCREEN_BLANK. Client: the accessibility service hangs a full-screen TYPE_ACCESSIBILITY_OVERLAY on itself. block=true makes it transparent and touchable so it swallows physical taps; black=true paints it black. Both are absolute states, so repeated commands are idempotent, and the handler clears the overlay when the screen sub-connection dies (Android has no "released when the calling thread exits" fallback). During a remote gesture the overlay is flipped to FLAG_NOT_TOUCHABLE, mirroring Windows' BlockInput(false) wrap; the flag change and the injection must be separated by a short delay or the gesture is swallowed. Server: CLIENT_CAP_PRIVACY_OVERLAY distinguishes new APKs from old ones; Linux/macOS and old APKs still get -32005 up front. Tool masks, instructions and mcp/tools.json follow the change. COMMAND_SCREEN_BLANK is implemented but not exposed as a tool: MediaProjection captures the overlay, so blacking out the screen blinds the operator too and cannot emulate Windows' WDA_EXCLUDEFROMCAPTURE. Verified on the Android 15 emulator: blocked taps are swallowed while remote clicks still land, the block re-arms after each gesture, the watchdog auto-unblocks at the deadline, a dropped sub-connection clears the overlay, and a black overlay shows up in both screencap and get_screenshot. The committed ghost.apk is stale; rebuild before release. Co-Authored-By: deepseek-flash |
||
|
|
445bd8bbb9 |
Feature: Support the remote terminal on Android hosts
The terminal tools were Windows/Linux only because the Android client had no COMMAND_SHELL handler: the server opened a shell sub-connection and the command was silently dropped, leaving exec_command to time out. This adds the PTY path on the client and the matching server gating. Client: COMMAND_SHELL on the main connection spawns a terminal thread that opens an authenticated sub-connection, sends TOKEN_TERMINAL_START and keeps the PTY alive until the server closes it. common/PTYHandler.h is reused, with a minSdk-21 openpty built from posix_openpt/grantpt and /system/bin/sh (mksh) as the shell. A new CLIENT_CAP_TERMINAL bit lets the server reject old APKs up front with -32005. Server: the tool masks, instructions and docs matrix accept APK. While verifying, exec_command left a second echoed copy of the command line in stdout: mksh forks with ECHO still on, so input written before its line editor takes over is echoed by the line discipline in addition to the shell's own redraw. The server now waits for the shell's first output (the prompt) before sending the command and sentinel, which restores the single-echo assumption the cleaning step relies on. terminal_exec on a ready session is unaffected. Verified on the Android 15 emulator: exec_command returns clean stdout and real exit codes for id/uname/df/ls (including a failing command), and terminal_open keeps cwd across commands, pipes work and close is clean. The committed ghost.apk is stale and must be rebuilt before release; a rebuilt APK was used for this testing. Co-Authored-By: deepseek-flash |
||
|
|
915c587964 |
Feature: Support remote chat on Android hosts
The chat tools were Windows-only because the Android client had no COMMAND_CHAT handler: the server asked for a chat sub-connection and the command was silently dropped, leaving chat_open to time out. This adds the full path on the client and the matching server gating. Client: COMMAND_CHAT on the main connection spawns a chat thread that opens an authenticated sub-connection, sends TOKEN_CHAT_START and waits for the server to bind it. Incoming GBK lines are decoded and dispatched to ChatSession, which drives an in-app chat activity; when accessibility is available the activity is pulled to the front, otherwise a tappable notification is posted. Outgoing text is strictly GBK encoded so emoji and unmappable characters are rejected instead of becoming '?'. When the server ends the session the activity closes itself, matching the Windows client's DestroyWindow, and a new session starts from an empty log. Server: a new CLIENT_CAP_REMOTE_CHAT bit lets the server tell new APKs from old ones. ResolveHost now accepts Windows and Android hosts and rejects older APKs up front with -32005 rather than waiting for the command to expire. Tool platform masks, instructions, the docs matrix and mcp/tools.json follow the same change. Verified end to end on the Android 15 emulator: chat_open pops the activity, Chinese text from the server renders, a device reply reaches chat_wait, and ending the session from the master closes the activity instead of leaving a dead window. The committed ghost.apk is stale and must be rebuilt before release. Co-Authored-By: deepseek-flash |
||
|
|
29d7ee1a18 |
Feature: Support host uninstall on Android and expose uninstall_host
The server's "uninstall host" action is the single-byte TOKEN_UNINSTALL on the main connection. Windows clients have handled it for a long time; the Android client ignored it, so the menu entry silently did nothing there. Handle it in the Android client the only way Android allows: ask the system to uninstall the package. The client first replies with COMMAND_BYE and drops the connection, then opens the package installer's confirmation dialog through the accessibility service, or posts a notification that resumes the dialog from MainActivity when accessibility is off. The user still has to confirm on the device; if they cancel, the client has already left the server and must be reopened by hand. The manifest gains REQUEST_DELETE_PACKAGES, without which the installer activity refuses the request outright. New APKs advertise CLIENT_CAP_UNINSTALL so the server can reject old clients instead of sending a command nobody handles. The new MCP tool uninstall_host is gated on McpManage and McpReadonly=0, rejects Linux and macOS hosts, and offers the same wait=true verification as power_action, reporting went_offline once the host disappears from the online list. Verified on the emulator end to end: the tool returned went_offline after 200 ms, the system dialog appeared, confirming it removed the package. Co-Authored-By: deepseek-flash |
||
|
|
fd1b085279 |
Feature: Add remote lock screen for Android hosts
Implement the Ctrl+Alt+Del path on Android as a system lock screen. Android has no secure attention sequence, so COMMAND_SCREEN_CTRL_ALT_DEL is mapped in the client to GLOBAL_ACTION_LOCK_SCREEN and the server exposes it as the new MCP tool remote_lock_screen. The Android client resolves ControlService.lockScreen()/canLockScreen() in nativeInit and dispatches the command from the screen sub-connection through the usual attach/detach JNI path. The lock action is posted to the accessibility service handler because performGlobalAction must not run on the receiver thread. Login advertises the new CLIENT_CAP_LOCK_SCREEN bit only when the accessibility service is bound and the API level is 28+, so the server can reject old APKs and Windows/Linux hosts up front instead of reporting a lock that never happened. There is no receipt: a successful call only means the command was delivered, and the tool description tells the caller to confirm with get_screenshot. Server side adds the schema and handler next to remote_clipboard, audits the call, and gates it on client type plus the capability bit. Since it is the only APK-only tool, check_platform_table.py gains a documented exemption from the Windows baseline invariant, and make_tools_json.py learns an APK-only platform set. tools.json is regenerated from the live server and the plan document now counts 62 tools. Co-Authored-By: deepseek-flash |
||
|
|
0a0b709e8a |
Feature: Support remote clipboard and input on Android hosts
The MCP remote-control family (remote_open / remote_close / remote_mouse / remote_keyboard / remote_clipboard) previously stopped at Windows and Linux, even though the Android client already translated injected events into touch gestures and global actions. This wires up the missing piece and opens the five tools to APK hosts. Client side, the screen sub-connection now handles COMMAND_SCREEN_SET_CLIPBOARD: the native handler forwards the raw GBK bytes to CaptureService.onClipboardText, which decodes them, strips the trailing NUL that the human clipboard-sync senders append, and writes the system clipboard from a background app (allowed on Android 10+, only reads are restricted). Pasting cannot be a Ctrl+V combo, so VK_V is mapped to ACTION_PASTE on the focused node, falling back to the accessibility focus and then to the first editable node, since WebView and self-drawn inputs often do not flag FOCUS_INPUT or isEditable. A new capability bit CLIENT_CAP_SCREEN_CLIPBOARD (0x0800) tells the server the APK understands both paths, so builds without the handler get an explicit -32005 instead of a package that is silently dropped. action=type is rejected outright for APK hosts - physical key events would only turn 'v' into a paste - and the error points at remote_clipboard + key=V. The platform table, server instructions and the five tool descriptions now describe Android support, which keys actually act as global actions (ESC, HOME, APPS, PRINTSCREEN), and that injection needs the accessibility service while capture needs an authorized MediaProjection session, both without receipts. Verified on the API 35 emulator against a Debug server: ASCII and GBK Chinese text round-tripped through the clipboard into a focused Settings search field, type / expect_hwnd / window_hwnd returned -32005, and PRINTSCREEN produced a saved screenshot. check_platform_table.py and verify_tools.py --offline pass. Co-Authored-By: deepseek-flash |
||
|
|
f29150d20e |
Feature: Add screen preview to the Android client
get_screenshot and the server's online-list and hover thumbnails were gated to Windows and X11 Linux. The Android client now answers COMMAND_SCREEN_PREVIEW_REQ on the main connection with TOKEN_SCREEN_PREVIEW_RSP plus a JPEG, and advertises CLIENT_CAP_SCREEN_PREVIEW at login only when it can really capture. Capture goes through AccessibilityService.takeScreenshot (API 30+), enabled by canTakeScreenshot in the service config. MediaProjection is deliberately not reused: its VirtualDisplay already feeds the H.264 encoder and Android 14+ refuses a second one from the same projection, while accessibility capture runs alongside the live stream. The JPEG is scaled to maxWidth and encoded in Kotlin; native only copies the bytes and drops a request while one is in flight. Only mode=screen is supported, and capabilities are re-evaluated on every login so turning the service off withdraws the capability. Also fix two review findings: a screenshot callback that lands after the 3 s timeout now recycles its bitmap instead of leaking it, and the server's failure hint lists the possible Android causes instead of asserting disabled accessibility. Verified on the emulator against a local debug server: get_screenshot returns 640x1422 and 720x1600 JPEGs, the 120-wide list thumbnail works, and both MCP tool-table checkers pass with apk added for get_screenshot, get_activity_history and get_client_log. Co-Authored-By: deepseek-flash |
||
|
|
b2c0d81396 |
Feature: Add window info RPC for foreground checks and window activation
Add a capability-gated main-connection RPC (COMMAND_MCP_WINDOW_INFO=179 / TOKEN_MCP_WINDOW_INFO=180) that returns visible top-level windows with UTF-8 titles, screen rectangles, foreground flag and z-order, so MCP can verify focus before injecting input. New tool get_foreground_window; host_list_windows switches to the richer fields; host_control_window and operate_window gain action=activate; remote_keyboard and remote_mouse gain expect_hwnd, require_foreground and report_foreground. Old clients keep the previous behavior: the RPC is refused up front via CLIENT_CAP_WINDOW_INFO. The client caps a response at 96 windows (about 101 KB, under MAX_SEND_BUFFER). Signal the cut-off in the mode byte (0x80) instead of silently dropping windows: host_list_windows now reports truncated=true and says the list is incomplete, and activate verification reports state=unknown with reason=list_truncated rather than claiming the window is gone. expect_hwnd is checked only after the session is validated, so an invalid session_id no longer burns a window RPC and no longer reports a window error for a session that does not exist. Nullable output fields (foreground, window) are declared as ["object","null"] so strict clients do not reject responses. Co-Authored-By: deepseek-flash |
||
|
|
356aa27108 |
Feature: Reject MCP tool calls the target client cannot handle
tools/list 不带参数,无法按目标设备过滤工具列表。60 个工具里有 33 个只在 Windows 客户端上有实现,模型无从得知,只能下发命令后白等超时。 DispatchSingleHostTool 开头按 kToolPlatforms 表核对目标主机平台,不匹配立刻 返回 -32005;批量路径经由同一函数自动覆盖。主机不存在或离线时放行,把 -32002 留给工具自己报,避免同一个问题出现两种错误码。 支持面是刻意设计的阶梯:win 60/60、lnx 27、mac 16、apk 10。60 个工具全部含 win,所以 Windows 客户端永远不会被这道门禁拒绝,行为与改动前完全一致。 平台表按服务端实际放行来标,不按客户端能力标。两处会分叉的地方: - exec_command / terminal_* / kill_process:macOS 客户端有实现(main.mm:753 的 ShellworkingThread、SystemManager 的 KILLPROCESS),但服务端刻意拦着 (McpServer.cpp:3652「macOS 本轮不开放」、6870),故标 win+lnx。 - list_windows:Linux/macOS 的 SystemManager 都有 COMMAND_WSLIST 分支,但 BuildListWindows 把命令发在主连接上,而这两个平台的主循环只处理 COMMAND_SYSTEM (起子连接),WSLIST 只在子连接里才被处理,走不到。实测 Linux 上直接超时 -32001。那两个 case 是给 SystemDlg 用的,故标 win。 copy_files_c2c 是唯一的非对称工具(源端必须 Windows,目标端可 win/lnx/mac, 反向不支持),用 roles 按参数名分别声明,C++ 侧 CollectTargetReqs 对应处理。 配套 mcp/tools.json(60 个工具,version 2)与 tools/mcp_tools/ 下的抓取、生成、 比对、校验脚本。check_platform_table.py 校验 C++ 表、tools.json、instructions 三方一致,并把「Windows 全量」做成不变式。 已在 A6(Windows 11)与 Ubuntu24(Linux)上实测:Windows 侧 14/14 通过且无误拒, Linux 侧 11 项拒绝全部命中门禁,copy_files_c2c win->lnx 成功、lnx->win 被拒。 开关 McpPlatformGuard(默认 1),关闭只失去提前拦截,不影响任何执行路径。 |
||
|
|
1c832ff7f2 |
Feature: Add MCP cancellation, progress notifications and dry runs
P11 closes the plan's last batch, entirely on the server side: no client change is required for any of it. Long-running tools now expose a cancellation flag that every wait loop checks in 500 ms slices. A cancelled tool returns through the cleanup branch it already had (late first packets still get absorbed, sub connections still get closed, pending entries still get deleted), so there is no second cleanup path to keep in sync, and the host is released within a second. Three sources set the flag: notifications/cancelled (matched on the exact JSON-RPC id, with a fallback to the display form so a request_id copied out of the session list round-trips), mcp_cancel_session (by request_id, host id or session id), and a failed SSE write when the client has gone away. The dispatcher rewrites a cancel-induced failure to -32800 so callers do not mistake it for a host fault. tools/call requests carrying _meta.progressToken are answered as text/event-stream on the same POST: the tool runs on its own thread and both the notifications/progress events and the final JSON-RPC response go out through one chunked provider. Requests without a token are unchanged. download_file reports files done vs. expected, upload_file reports bytes sent vs. total. Two new local-only tools, mcp_list_sessions and mcp_cancel_session, list and cancel in-flight sessions (calls, pending one-shots, terminal, remote control, file transfer, chat). They are deliberately not gated by the feature switches or readonly mode, since that is the only way out of a stuck request. The -32003 busy message now names the occupier: "Device busy: ... (tool=download_file, elapsed=12s)". power_action gains wait/wait_ms to verify the result against the main connection's presence (went_offline, came_back, elapsed_ms); not seeing the transition is reported as ok, not as a failure. kill_process, operate_window and the four host_control_* tools accept dry_run, which validates the target against one fresh snapshot and returns what would be affected without sending anything. Validation rejections now use -32012; -32008 stays as documented historical alias. One implementation note worth recording: the per-thread call context uses FLS, not thread_local. Any static thread_local in this server crashes the process about a second after start, because some threads come from shellcode/MemoryModule images that never ran PE TLS initialisation. Verified against a real Debug server and a real Windows client over MCP: the smoke test now passes 104 checks with 0 failures, including cancellation releasing the host in under a second, the cancelled request returning -32800, notifications/cancelled returning 202, 63 progress events with exact byte counts for a 64 MB upload, and a cleanly terminated SSE stream. Not exercised in this pass: power_action's wait=true path (it would take the test host down) and the -32006 readonly/switch-off section (every switch was on during the run). Co-Authored-By: deepseek-flash |
||
|
|
3d388f8e45 |
Feature: Add remote_block_input to MCP remote control
A GUI automation run is interrupted the moment a person at the host touches the mouse, and locking the host's keyboard and mouse is the standard answer. The client already implements COMMAND_SCREEN_BLOCK_INPUT and the MFC screen window already sends it, so the tool needs no client change: it puts the same two-byte packet [COMMAND_SCREEN_BLOCK_INPUT][0|1] on the hidden screen sub-connection that remote_open established. The gate is remote_open's own - McpRemoteControl=1 and McpReadonly=0 - because the ability to lock someone out deserves exactly the same door as the ability to drive their desktop, and Linux and Android clients are refused with -32005: they have no branch for this packet at all and would silently ignore it. Locking a host out is only acceptable if every way back in is closed, so the session now carries its own lock state and three paths release it. block=false releases it directly. block=true always carries a deadline (max_block_ms, default 300000, clamped to 1000..3600000) and a watchdog thread started with the MCP server sweeps every 500ms for expired locks, sends the release and records it; the next remote_block_input reports it as auto_unblocked so the model learns that its lock lapsed instead of assuming it is still held. remote_close and the idle reclaim both send block=false before they tear the sub-connection down, because once that connection is gone the server has no channel left and a live client process stays blocked - all three release paths write an audit line, since none of them is something a person can see. The send happens outside the MCP lock, and the watchdog takes the session's busy flag while it releases, so it cannot interleave with an injection or race a close; a failed send keeps the lock marked and retries five seconds later rather than hammering a dead connection every tick, and a session that dies under the watchdog is erased the same way an injection would erase it. Verified against a real Windows client over MCP on a Debug x64 build: block then unblock, block with a 1.5-2s deadline left to expire and reported back as auto_unblocked with the watchdog audit entry at the expected second, block with a ten-minute deadline then remote_close with "(input lock forced off)" in the audit, and the five refusals (missing block, non-boolean block, max_block_ms out of range, unknown session, offline host). The smoke test gained [4d] for those refusals; 54 checks pass. Two limits are worth stating plainly. BlockInput only works when the calling process is elevated (or UIAccess), the client ignores the API's return value, so on an ordinary client the lock silently does nothing and the tool can only promise that the command was delivered - both the tool description and the server instructions say so. And the idle-reclaim release, like the two verified paths, is only exercised at a 300s idle timeout; it shares the same send helper and was reviewed rather than timed out live. Bin/settings.ini was restored byte-exact and the Release instance was left alone. Co-Authored-By: deepseek-flash |
||
|
|
27b5431130 |
Feature: Add the file-write tools behind their own switch
P10-B adds delete_path, move_path and create_folder, and the switch they need. Reusing McpFileTransfer would have been wrong: uploading overwrites a file, deleting destroys a tree, and an operator who wants one may well not want the other. McpFileWrite is therefore a separate key, off by default, wired the way every other switch is - read from settings.ini at startup, reported as get_server_status.switches.file_write, and exposed with its own checkbox in the MCP settings dialog. The tools additionally require readonly to be off: a read-only server has no business modifying a host. No client change was needed: COMMAND_DELETE_FILE, COMMAND_DELETE_DIRECTORY, COMMAND_CREATE_FOLDER and COMMAND_RENAME_FILE already existed in the Windows file manager. The chain reuses the file-manager sub-connection and adds a phase - the main connection asks for the drive list, the sub-connection's first packet hands it to OnDriveList which sends the write command, and the client's TOKEN_*_FINISH ack moves the pending request from phase 0 to 1, where the tool sends COMMAND_LIST_FILES for the parent so the ack and the confirmation share one timeout. Only the IO thread touches the sub-connection; the tool thread just waits on the shared condvar, which keeps the sub-connection's lifetime out of its reach. The confirmation exists because the client answers unconditionally: it ignores the return values of DeleteFile, MoveFile and MakeSureDirectoryPathExists and there is no error token. ok therefore only means "the command arrived", and verified is the real answer - true when the parent listing shows the expected post-state, false when it shows the opposite, null when the listing itself did not come back (timeout or disconnect), since reporting that as false would be a lie. Linux, macOS and Android hosts are refused outright: those clients decode these paths as GBK while the server speaks UTF-8 to them for list_files, so a non-ASCII path would corrupt rather than fail. Paths are validated as UTF-8, converted to GBK exactly once, and reject %VAR%, relative paths, drive and share roots, and - for delete_path and for move_path's source - the Windows, Program Files, ProgramData, $Recycle.Bin and System Volume Information trees. create_folder gets no system-directory ban, because writing there is the same permission upload_file already has. Two defects surfaced while exercising this against a real client: create_folder's wire path has to end with a backslash, because MakeSureDirectoryPathExists treats a non-terminated last component as a file name and only creates the intermediate levels; and ResolveFileOpVerified(LookupInListing(...), isDir, ...) read isDir before the lookup ever set it - argument evaluation order is unspecified and MSVC evaluates right to left - so every create_folder reported verified=false even when the directory existed. Verified on a Debug x64 build against a Windows client over MCP: nested create_folder, upload, rename, delete, delete a tree, the failure shapes (delete with recursive=false on a directory, create_folder onto an existing file name, delete of a non-existent target) and every refusal (drive root, C:\Windows\Temp, C:\Users, %TEMP%, relative, empty, over-long, from==to, offline host, non-numeric id), with a Chinese-named directory created, renamed and deleted end to end; no file-manager window appeared in the server process during any of it. The smoke test gained section [4c], opt-in through -FileWriteDir, which asserts verified and never ok and walks a sandbox through the same path; two runs passed 69 checks with the sandbox removed afterwards. Bin/settings.ini was restored byte-exact and the Release instance was left alone. Co-Authored-By: deepseek-flash |
||
|
|
bb55f52b29 |
Feature: Add instructions, capability reporting and filters to MCP
P10-A is the first batch of the next MCP phase: the server now tells a model
how it expects to be driven instead of leaving the rules to be inferred from
per-tool descriptions.
initialize carries the instructions field that MCP 2025-06-18 defines, in
five sections: where host ids come from, which switch hides which tool, what
each error code implies for the next action, typical orchestrations, and
safety. Every statement was checked against the implementation rather than
derived from the tool descriptions, and four did not survive: the host_*
family only supports name/pid filters in part (named per tool now), read-only
mode does not hide host_set_remark and host_set_watch_level (they touch only
the local host store and are refused with -32006 when called), -32003 is one
lock keyed by device id rather than one per capability, and there is no
"screen stream" tool to pick instead of get_screenshot. The string is built
from u8 literals because the project compiles with /execution-charset:.936.
get_host_detail returns a capabilities object (platform, screenPreview,
remoteControl, fileTransferV2, c2cPaths, registry), each flag tied to a gate
that already exists in the dispatcher, so a model no longer discovers a
host's limits by calling and reading -32005. remoteControl mirrors the two
conditions remote_* actually applies (Windows or Linux, and not
multi-monitor), not the preview bit. Making the flags honest exposed
list_registry rejecting only LNX and MAC, so an APK host passed and then
waited out the timeout; it now uses the same is-Windows predicate the bit is
computed from. That predicate existed as identical local copies in
McpMachine.cpp and McpChat.cpp and now lives in McpInternal.
list_processes and list_windows take an optional name and pid, filtered on
the server before the page window is applied, with total counting matches
and totalAll still reporting the unfiltered count. The two tools get their
own schema builder so list_services does not advertise a filter it would
silently ignore, and the pid descriptions differ because a pid means the
process itself for one tool and its owner for the other.
get_keylog gains set_offline and clear, the two commands the client already
implements on the keyboard sub-connection but nothing on the MCP side ever
sent, plus offlineRecord from the ready packet's byte 1. Both writes are
gated on McpReadonly=0 and validated before COMMAND_KEYBOARD goes out, so a
refused call cannot leave a running keyboard session behind; the tool's
annotations change from read-only to writable and destructive, and the
server instructions gain the matching sentence.
Four capabilities are exposed twice through two different sub-connection
chains, so each pair's descriptions now say which one to reach for first and
when the other is the only option: the host_* trio carries more fields and
filters on the server, the list_* trio is the McpMachine=0 fallback, and
operate_window is the only one that can close a window. No tool was merged:
the split is what lets McpMachine be switched off on its own. search_file's
description drops the stale MAX_PATH-overflow warning (fixed in
|
||
|
|
c271911fcb |
Feature: Add copy_files_c2c to transfer files between clients directly
Adds the 54th MCP tool, copy_files_c2c, which asks the source client to push a set of files straight to another online client. Previously the only way to move files between two clients was to download them to the server machine and upload them again, so the data crossed the wire twice and needed free disk on the server. With C2C the server only relays packets and never parses or stores the payload. Protocol. Two new ids in common/commands.h: COMMAND_C2C_SEND_FILES (174), carrying the destination directory followed by a NUL-separated path list, and TOKEN_C2C_SEND_RESULT (175), a fixed 45-byte result packet sent back over the same sub-connection the transfer used. Clients advertise support with the LOGIN_INFOR capability bit CLIENT_CAP_C2C_PATHS (0x0008), exposed server side as Context::SupportsC2CPaths(). File transfer V2 already knew how to push to dstClientID, so the destination side needed no change at all; existing V2 receive paths keep working, and older clients simply do not advertise the bit and are rejected with a clear error instead of hanging. Client. OnC2CSendFiles expands environment variables, collects directories recursively, opens its own sub-connection (with sub-connection auth), and runs FileBatchTransferWorkerV2 in a detached thread, then reports the result on that same connection. Every early exit now reports a structured failure rather than staying silent. The worker needs the process-level InitFileUpload call that the receive path performs lazily, and a source client may never have opened a file manager or received an upload, so the initialization is factored into EnsureV2TransferInited() and called from both paths. Server. copy_files_c2c mirrors the existing upload tool: gates, id parsing, capability checks, path validation, then a session keyed on the source host that WaitC2CDone polls until the result packet arrives. The C2C relay branch used to attach a CDlgFileSend progress dialog to the source's sub-connection; NotifyProc then routed every later packet on that connection to the dialog, which parsed the result packet as a V1 file chunk and dropped it, so the tool always timed out even though the files had arrived. IsC2CTransferPending() now suppresses dialog creation for MCP-driven transfers only, leaving manual Ctrl+V transfers untouched. Errors map to -32001 for a source-side transfer failure, -32010 for access denied, -32009 for a missing path and -32002 for an offline destination. Both hosts must be Windows: the destination is restricted too, because the source embeds dst_dir in each V2 chunk name as an absolute Windows path and only the Windows receiver applies it literally. A Linux or macOS receiver never gets a target directory here (the clipboard flow sets one with COMMAND_C2C_PREPARE, which this path does not send), so it would strip the drive letter and write under / instead. Supporting that needs a prepare packet carrying an explicit path plus an upgrade of those clients, so for now the tool rejects them with -32005 rather than writing somewhere wrong. Verified end to end on the Debug build: three files including a non-ASCII name, an empty directory and a 3 MB binary arrived with matching SHA-256 in 3.2 s, a non-existent destination directory was created, and the negative cases returned the documented codes. mcp_smoke.ps1 passes 41/41 with two new assertions and 27/27 in -SkipHost mode. Design notes, error table and limitations are in docs/Mcp_FileTransfer_Design.md section 14 and docs/Mcp_Design.md section 15. Co-Authored-By: deepseek-flash |
||
|
|
27d40f5678 |
Fix: Expose tool payloads in the content text block
List and query tools put only a human summary in content and kept the real payload in structuredContent. Clients that hand just content to the model (WorkBuddy does) therefore never saw a host id: search_hosts answered "1 host matched" and every follow-up tool was out of reach, and guessed ids 1-3100 could never work because host ids are 64-bit snowflake values. BuildResult now appends the serialized structuredContent as a second text block, as the MCP spec recommends for tools with structured output. The append sits in the single result choke point, so every tool in McpServer, McpMachine and McpChat gains it, including future ones; empty action results such as remote_mouse keep their bare "ok". Tools whose text already carries the payload expose only the difference: exec_command and terminal_exec add exit_code, get_keylog adds counters, get_activity_history exposes records without the duplicated raw dump, and the hosts file reader drops its copy of the content. The override rides in a $ key that BuildResult consumes and removes, so responses change only by the extra block. Verified on a Debug instance: list_all_hosts, search_hosts, get_server_status and list_registry return parseable JSON in content[1] with ids intact, and tools/mcp_smoke.ps1 -SkipHost passes 27/27 with four new assertions. Co-Authored-By: deepseek-flash |
||
|
|
3b1626f415 |
Feature: Turn client failure reports into branchable MCP errors
Wire the four new failure tokens into MessageHandle and translate them in the tool builders, so the three previously silent failures now return a code an agent can branch on: list_services on a non-elevated client is -32010 in well under a second instead of a 20 s timeout, list_files and list_registry distinguish a missing path or key (-32009) from an empty one, and download_file no longer waits for the stall watchdog when the remote path does not exist. Access denied gets its own code, -32010, and the win32 code is kept in the message for every case. TakeRegError fills both registry wait flags in one packet because a failing client sends no subkey or value packet at all. OnDownloadError sets the session error without cancelling the file-manager sub-connection, which stays owned by ClearFileTransfer. ListRemoteTopLevelNames treats "not found" as "directory not created yet" so uploading into a fresh directory keeps working now that the client reports it. Also converge the error codes that were overloaded: a multi-monitor host is a capability limit, not a parameter mistake, so remote_open reports -32005 instead of -32008. The tool wait timeout moves from a compile-time constant to McpToolTimeoutMs in settings.ini (default 20000, clamped to 1-600 s), which makes it possible to tune without a rebuild. The smoke test gains the three reproducible negatives. Co-Authored-By: deepseek-flash |
||
|
|
cc3e11d30d |
Improve: Cover protocol and policy error codes in the MCP smoke test
The suite only exercised the happy paths, so a regression in the negative paths -- unknown method, unknown tool, a missing required argument, an oversized batch or a policy refusal -- would have gone unnoticed. Add those checks alongside a section that reads the switches from get_server_status, picks a feature that is actually turned off and asserts its tool answers -32006; with every switch on it skips rather than guessing, and with -AuditLog it also asserts the refusal shows up as CALL_DENIED. Document the full error-code table in section 12, including which codes mean "retry with different arguments" and which mean "this is a policy decision, retrying is pointless", and note that a batch call never fails as a whole for per-host errors. Verified against the Debug instance with McpKeylog temporarily set to zero: get_keylog refused with -32006 and the audit log gained the matching CALL_DENIED line, 34 checks passing; with the switch restored 32 pass, and 37 with a search path. Co-Authored-By: deepseek-flash |
||
|
|
7c5bf3e73d |
Feature: Filter search_file results by size, mtime and slim fields
Add four optional arguments to search_file that narrow the match list after the client has returned it: min_size and max_size in bytes, modified_after in either YYYY-MM-DD or YYYY-MM-DD HH:MM:SS form, and name_only to keep just path and name per entry. The client still runs the same name-matched recursive search, so these filters cut the response size and the reasoning work on our side, not the scan time on the host -- the summary sentence says so and totalAll keeps the pre-filter count so a too-narrow condition is obvious. Invalid values (negative sizes, an inverted range, a malformed timestamp) are rejected with -32602 before any packet goes out. The timestamp validator only accepted the ten-character form because its shape loop never allowed the two colons, which rejected every nineteen-character value; the loop now whitelists them. Smoke tests cover the unchanged totalAll under a size filter, the slim name_only payload and a rejected timestamp, for 33 checks when the search path is supplied. Design notes are in section 11.4. Co-Authored-By: deepseek-flash |
||
|
|
929065e7be |
Feature: Filter the MCP host lists by group, OS and client type
list_online_hosts and list_all_hosts could only page, so finding "the hosts in this group" or "the Linux ones" meant walking pages or falling back to search_hosts, which has no client type filter and treats online and offline hosts differently. Both tools now accept group, os and client_type. They are substring matches, combined with AND and case-insensitive over ASCII, matching the semantics search_hosts already used, and the shared implementation keeps the two tools from drifting apart. The filter runs before paging, so total counts the matches and totalAll keeps the pre-filter size, and the summary switches to the "matches" wording with a trailing line naming the unfiltered size - a mistyped condition then shows up immediately instead of silently returning nothing. Offline hosts carry no clientType in their stored record, so a client_type filter never matches them; the tool description says so and points at os for that case. Verified on the Debug instance: client_type=exe matches the single EXE host while client_type=lnx matches nothing and still reports totalAll=1; list_all_hosts with group=default matches all 3161 recorded hosts and splits them into 1 online and 3160 offline, and os=zzz matches none. The smoke test gains two checks and passes 28/28. Co-Authored-By: deepseek-flash |
||
|
|
5ce453797c |
Feature: Add a get_server_status tool to the MCP server
Answering "did my rebuild actually take effect" or "why was that tool refused" used to mean reading list_online_hosts and get_audit_log and inferring the rest. A single read-only call now returns the server's own state: brand and version, the exe link time (taken from the file's last write, which unlike __DATE__ reflects the whole build), how long the process has been up, the address and port MCP is actually listening on, every MCP switch, online and offline host counts, the number of sessions in flight per family, and how many source IPs are currently blocked. The bind and port are recorded by Start rather than re-read from the configuration, so the tool reports what the listener really got. Session counters take each subsystem's lock in turn and release it immediately, so the numbers are an estimate of one moment, not a consistent snapshot. Nothing sensitive is exposed: the token and the command whitelist never appear, and the switches are reported as booleans only. The tool is registered as read-only, non-destructive, idempotent and server-local, which brings tools/list to 53 entries. The smoke test gains three checks for it, including one that fails if a token-shaped field ever appears in the output. Verified on the Debug instance: the summary line reports version, uptime and listener correctly, sessions.terminal goes 0 to 1 while a persistent terminal is open and back to 0 after terminal_close, and the smoke test passes 26/26. Co-Authored-By: deepseek-flash |
||
|
|
ae37e812ee |
Feature: Page and filter the MCP audit log tool
get_audit_log used to hand back every entry in the main window's message mirror on each call. The mirror is a 1000-entry ring buffer, so a busy server returned a lot of text the caller never looked at, and there was no way to ask for just the MCP entries or just the ones mentioning a particular subject. The tool now takes the same limit/offset pair as the other list tools and two optional substring filters: type matches the type column and keyword matches the message text, both case-insensitive. Filtering runs before paging, so total counts the matches while totalAll keeps the size of the whole mirror, and the summary line adds the pre-filter count when a filter is in use so a typo is easy to spot. The GBK to UTF-8 conversion moved into the filter loop so the comparison happens between strings in the same encoding, still under the dialog's critical section. The output schema gains total/totalAll/offset/limit/hasMore, and the smoke test now asserts the paging fields, the limit bound and an empty result for a filter that cannot match, bringing it to 23 checks. Verified on the Debug instance: limit=2 returns entries 1-2 then 3-4 with hasMore set, offset=999 reports the out-of-range text, type=mcp matches the "MCP命令执行" entry, type=提示 matches both hint types and keyword=监听 matches the listening-port line. Co-Authored-By: deepseek-flash |
||
|
|
abe7e99a77 |
Feature: Harden the MCP transport and add non-destructive host tools
Cap MCP request bodies at 2 MB, set a 15 s idle read timeout and a 60 s
write timeout, and give httplib an explicit 32-thread pool: long tools
such as chat_wait hold a worker for up to 55 s, and the default pool of
8 to 15 left other callers queued.
Every tool returned by tools/list now carries readOnlyHint,
destructiveHint, idempotentHint and openWorldHint. The table is
per-tool; anything not registered defaults to destructive so a new
write tool cannot silently look harmless. Read-only tools are forced
to destructiveHint=false.
get_host_detail appends the technical fields the host store already
holds (cpuCores, memoryGB, systemBits, programBits, user, isAdmin,
installTime, expiredDate, monitors, exeVersion, clientPath, pid,
publicIp). The JSON writer is pinned to 15 significant digits so
memoryGB renders as 27.8 instead of 27.800000000000001. list hosts
output is unchanged.
Add host_set_remark and host_set_watch_level. Both validate the host
against the store, are refused in readonly mode, repaint the affected
row, and are audited without recording the value written.
Add search_file, which drives the client file-manager search channel
and returns paged results. The root path is normalized before sending:
the client enumerates with wsprintf("%s\*.*"), so a trailing
backslash produced C:\*.* and FindFirstFile failed, returning zero
hits instantly. The client only searches recursively, so non-recursive
requests are filtered server-side by comparing the parent directory.
Collection stops at 5000 records or 4 MB and the client is told to
stop; on timeout the partial results are returned. Search result
packets are never closed early: the sub-connection stays open until
TOKEN_SEARCH_FILE_FINISH, because the client's search thread holds a
pointer into the file manager object.
Batch reads report a skipped counter and are bounded to 50 ids and a
50 s deadline.
Add tools/mcp_smoke.ps1, a read-only end-to-end smoke test that asserts
auth rejection, the payload limit, annotation coverage, paging fields
and the main read tools. It takes the token from the environment and
never prints it.
Document all of the above in docs/Mcp_Design.md section 10, including
the client-side MAX_PATH stack overflow that a deep search_file call
can trigger, together with the workaround of searching shallow
directories.
Co-Authored-By: deepseek-flash
|