* fix(runner): align SDK pin and workspace-aware integration fixtures
* fix(ci): format sources and resolve current migration head
* test(persistence): align standalone migration fixtures with current models
* test(web): align smoke fixtures with current processor UI
---------
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
execute_func_call returns list[ContentElement] for MCP tools, but the
runner assigned that list directly to the tool-message content. The
OpenAI chat-completions spec requires tool-message content to be a
string, so OpenAI-compatible endpoints return HTTP 500 when the raw
list is sent.
Serialize the list to a string before building the tool message, using
ContentElement.__str__ which returns the text payload for text elements
and a human-readable placeholder for images and files. Fixes#2457.
SessionManager clears image_base64 on past turns to save memory, and
exclude_none serialization drops the hollowed field entirely, so a
replayed history part can arrive as {'type': 'image_base64'} with no
payload. The converter accessed the missing key unconditionally and
raised KeyError on every turn after an image was sent.
Prefer the base64 payload when present, fall back to an image_url that
survived on the same element, and drop hollow parts otherwise (same
strategy as the existing file-part handling). Fixes#2469.
Previously, MCP tool calls via call_tool() had no timeout, so a hung MCP
server would block the entire session indefinitely (exacerbated by
concurrency.session=1). This wraps the call in asyncio.timeout(30) and
raises an Exception on expiry, letting the LLM recover gracefully.
Closes#2339
When the LLM (e.g. MiniMax-M3) returns multiple rounds after tool calls, the _StreamAccumulator was initialized with initial_content=first_content, causing every subsequent round to repeat the entire first message. Remove the re-seeding so each round starts with a clean accumulator.
Add regression test verifying multi-round tool call content is not duplicated.
Drop the PluginToolLoader.get_tool() override that returned a raw
ComponentManifest, so every loader's get_tool() now returns a uniform
resource_tool.LLMTool (PluginToolLoader.get_tools() already did this
conversion). This removes the only source of tool-shape heterogeneity.
- ToolManager.get_tool_schema(): drop the ComponentManifest-vs-LLMTool branch
- ToolManager.get_tool_detail(): new host-level shape {name, description,
human_desc, parameters}
- handler.py GET_TOOL_DETAIL: call tool_mgr.get_tool_detail(); delete the
handler-local _build_tool_detail + _i18n_to_dict/_i18n_to_text adapters and
the litellm TODO
- ToolLookupResult is now just LLMTool
The dropped label/spec fields were not consumed by any runner (local-agent
build_llm_tool and external harnesses use only name/description/parameters).
Expose skill tools (activate/register_skill/native exec) like native tools
instead of gating them behind the skill_authoring capability:
- toolmgr.get_all_tools drops include_skill_authoring; SkillToolLoader
self-gates on sandbox + skill_mgr
- preproc drops the include_skill_authoring branch; pipeline-bound skills
and the skills resource gate on skill_mgr presence
Persist activated skills into host.activated_skills conversation state so
they survive across runs (host writes at activate; last-write-wins); drop
the dead restore_activated_skills helper.
Prefill ToolResource.parameters host-side (tool_mgr.get_tool_schema) so
runners build LLM tools without per-tool get_tool_detail round-trips.
Align agent-runner-pluginization design docs to the all-tool model.
* fix(provider): strip think tags for MiniMax-M3 and other OpenAI-compatible models
MiniMax-M3 (and other OpenAI-compatible providers) emit chain-of-thought
reasoning directly in the content field wrapped in tags, instead
of using a separate reasoning_content field or the legacy CRETIRE_REASONING
markers. The existing remove_think logic only handled CRETIRE_* tags, so
think blocks leaked into user-visible output even when remove_think was enabled.
- Add _ThinkStripState: a stateful filter that correctly handles tags
split across streaming chunk boundaries.
- Add _strip_think classmethod with regex patterns for both and
CRETIRE_* tags.
- Wire think_state into invoke_llm_stream so deltas are filtered before
reaching the accumulator.
- Add remove_think safety net in _StreamAccumulator so the final message
from tool-call rounds also gets stripped.
- Fix remove_think resolution to use defensive nested .get() so
pipelines missing output.misc don't raise AttributeError.
* fix(litellmchat): add missing _CLOSE_TAG class attribute on _ThinkStripState
* fix(provider): handle think stripping across LiteLLM paths
---------
Co-authored-by: WangCham <651122857@qq.com>
* feat: Implement workflow form handling for paused workflows
- Added module-level storage for pending forms to manage state across sessions.
- Introduced functions to set, get, and clear pending forms with expiration handling.
- Enhanced DifyServiceAPIRunner to support resuming paused workflows via form actions.
- Implemented logic to yield human input requests and display appropriate messages.
- Updated workflow submission methods to handle paused states and resume actions.
- Ensured proper merging of pending form actions with user inputs for seamless interaction.
* feat: Add '_routed_by_rule' variable to form action in Lark and Telegram adapters
* feat: Enhance Lark and Telegram adapters with new form handling for paused workflows
* feat: Enhance TelegramAdapter to handle form action buttons and message threading
* feat: Improve TelegramAdapter message handling with enhanced error management and draft message support
* feat: Add the function for formatting human input text to support adapters without rich UI.
* feat(dingtalk): implement human input card support and card action handling
- Add a new module `card_callback.py` to handle card action button clicks from DingTalk.
- Introduce `DingTalkCardActionHandler` to process card action callbacks and extract parameters.
- Update `DingTalkAdapter` to manage card state and handle form input through a single card template.
- Add configuration for `human_input_card_template_id` in `dingtalk.yaml` to specify the template for human input.
- Create a new card template `dingtalk_human_input_card.json` for rendering human input prompts and buttons.
* feat(dingtalk): enhance human input card functionality with streaming support and active turn management
- Updated the DingTalk card template to enable streaming mode and multi-update configuration.
- Removed the obsolete delete_card method from DingTalkClient to streamline card management.
- Enhanced DingTalkAdapter to manage active turn cards and accumulated streaming text, ensuring a seamless user experience during human input prompts.
- Modified the create_message_card method to utilize existing active cards for resumed workflows, preventing duplication.
- Improved the _paint_form_on_card method to update existing cards with human input prompts and buttons dynamically.
- Updated the dingtalk_human_input_card.json template to reflect the new streaming capabilities and configuration options.
* feat(wecom): implement Dify human input pause handling with button interaction support
* feat(qqofficial): implement Dify human input button interaction handling and markdown keyboard support
* feat(qqofficial): implement one-click QR binding and enhance localization support
* feat(discord): implement Discord form view with button interactions for Dify actions
* fix(telegram): correct group chat type check and handle oversized callback data for Telegram actions
fix(difysvapi): ensure safe access to remove-think configuration in pipeline settings
* feat(dify): add support for chatflow app type and enhance human input handling
* feat(telegram): add action title feedback for user selections in Telegram messages
* feat(lark): enhance LarkAdapter to store form content for resume notices
* feat(dingtalk): update display formatting for card content with HTML line breaks
* feat(dingtalk): add feedback functionality to cards with 👍/👎 buttons
- Implemented feedback state management for cards, allowing users to provide feedback via thumbs up/down buttons.
- Enhanced card rendering to include feedback buttons when appropriate.
- Registered feedback listeners to handle feedback events and update card states accordingly.
- Updated the card template to support dynamic button rendering for feedback actions.
- Improved error handling and logging for feedback actions and card updates.
* fix: add Avatar component to dingtalk_human_input_card.json for enhanced user interaction
* feat(wecom): add optional source block to interactive template cards for enhanced branding
* feat(wecom): add functions for template card action extraction and update, enhance button interaction handling
* feat(qqofficial): synchronize passive-reply counter with inbound message sequence
* feat(qqofficial): add method to identify invisible form placeholder chunks in messages
* feat(dingtalk): add download link for human input card template and enhance dynamic form configuration
* feat(telegram): enhance message handling with group stream deletion and form placeholder detection
* Add unit tests for DingTalk, Lark, WeComBot, and Dify service API runners
- Implement tests for DingTalk adapter helper functions including form content cleaning, input extraction, and completed input lines.
- Create unit tests for Lark adapter helper functions focusing on input extraction and completed input lines.
- Add tests for WeComBot template card functionalities, including event extraction and payload building for human input.
- Enhance Dify service API runner tests to cover human input forms, including input collection, action handling, and form snapshot extraction.
* feat: Enhance Telegram and QQ Official adapters with select field handling and form action processing
- Added support for select fields in Telegram adapter, including option extraction and callback handling.
- Implemented form action processing for Telegram callbacks, improving user interaction feedback.
- Introduced new helper functions for building keyboards and resolving select button actions in QQ Official adapter.
- Enhanced DifyServiceAPIRunner to handle cumulative streaming responses and improve error handling during workflow resumes.
- Added unit tests for new functionalities in Telegram and QQ Official adapters, ensuring robust behavior for select fields and form actions.
* feat(lark): add functions for current input definitions and visible form content handling
feat(qqofficial): update fallback text handling for non-streaming scenarios
feat(difysvapi): enhance form content processing for interactive fields and actions
test: add unit tests for Lark and QQ Official adapter functionalities
* Add tests for DingTalk adapter content processing and markdown formatting
- Updated the assertion in `test_dingtalk_completed_input_lines_include_text_and_select_values` to remove unnecessary markdown formatting.
- Added new tests to verify that `_dingtalk_clean_form_content` maintains the order of prompts and completed values in various scenarios.
- Introduced `test_dingtalk_card_markdown_preserves_internal_line_breaks` to ensure internal line breaks are correctly converted to HTML line breaks.
* feat: Refactor input handling and feedback messages across multiple adapters
* feat: Update the human-computer interaction template cards, and optimize the prompt information and content display.
* feat: Refactor pending form handling to isolate by bot and pipeline
* feat: Enhance error handling and caching for Dify and WeCom interactions
* feat: Enhance select input handling and validation in Dify API runner and Telegram adapter
* feat: Add missing completed input lines handling in DingTalk adapter
* feat: Add pipeline_uuid handling across multiple adapters and update related tests
Operators can now set a global default memory limit for all stdio MCP
servers in config.yaml or via environment variable:
config.yaml:
box:
default_memory_mb: 2048 # default: 1536
env:
BOX__DEFAULT_MEMORY_MB=2048
The default is raised from 1024 to 1536 MB — a safer floor for
Node.js V8 + WASM (undici llhttp) under nsjail cgroup limits.
Individual MCP servers can still override via their own box.memory_mb.
Previously the fallback was hardcoded to 1024 MB, causing OOM kills
(return_code=137) on node/npx MCP servers that need more RAM.
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
All MCPs share one Box session (mcp-shared). When session memory_mb differed
by command type (512 for python, 1024 for node), the second MCP to call
create_session raised BoxSessionConflictError. Fix: always use 1024 MB for
the shared session so python and node MCPs coexist without conflict.
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
Node.js MCP servers (npx/bunx) were being OOM-killed (return_code=137) by the
default 512MB nsjail cgroup_mem_max. Node V8 reserves large virtual address
space and instantiates WebAssembly modules (undici llhttp) on startup, easily
exceeding 512MB resident. This caused every node-based MCP (memory,
sequential-thinking, filesystem, weather, docker, excel) to crash-loop.
Fix: when the stdio command is npx/bunx/pnpm, default memory_mb to 1024 unless
the operator explicitly set a value. Python/uvx servers keep the 512MB default.
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
_ColdStartRetry was caught in _lifecycle_loop_with_retry which set
_preserve_managed_process = True — but by then the finally block inside
_lifecycle_loop had already run and called _cleanup_box_stdio_session(),
stopping the live managed process (return_code=143 SIGTERM). The cold-start
retry then restarted a fresh process, eliminating the warm-up advantage.
Fix: add an explicit except _ColdStartRetry in _lifecycle_loop that sets
_preserve_managed_process = True before re-raising. The finally block then
sees the flag and skips stop_managed_process, leaving the live process
untouched for the next handshake attempt.
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
The previous _TransferredStack approach broke anyio lexical context:
websocket_client/ClientSession use anyio task groups whose cancel scope is
bound to the frame that entered them. Deferring their aclose via a transferred
exit stack left the underlying memory streams closed once initialize() returned,
so the very next request (refresh -> list_tools) failed with Connection closed.
New design:
- Attach on the owner exit stack (same task as the serve loop, lexically intact)
- A cold-starting process makes initialize() fail; signal _ColdStartRetry up to
the outer retry loop, which reuses the live process without consuming retry budget
- _lifecycle_loop_with_retry handles _ColdStartRetry like _TransportReconnect:
preserves process, no fatal budget, backs off 2s and retries
- Two new unit tests: cold-start raises _ColdStartRetry (not fatal) when process
is alive; raises fatal error when process has actually exited
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
A node/npx stdio MCP server (e.g. firecrawl-mcp via npx -y) failed on first
connect with Connection closed / Failed after 4 attempts, even though the
process was fine. An npx cold start downloads+installs the package before the
server can answer the MCP handshake (measured ~27s for a simple official
server; longer for heavier ones). The old code attached the WS and called
session.initialize() the instant the process was started, so the handshake ran
before the process could answer and failed; the outer lifecycle retry then
rebuilt the process, churning it in a loop.
Verified decisively: attaching + initialize() against a mid-cold-start process
times out on attempt 1 (process still installing) but SUCCEEDS at t+0.6s on
attempt 2 once the process is ready. So the fix is to retry the handshake in
place, not to rebuild the process.
Changes (mcp_stdio.initialize):
- Start the managed process ONCE, then loop attach WS -> ClientSession ->
initialize() within the startup_timeout budget, tearing down each failed
attempt cleanly, until the handshake succeeds or the budget elapses. A
successful transport/session is transferred into the owner exit stack via a
small _TransferredStack adapter.
- Bound each attempt with asyncio.wait_for(initialize, _HANDSHAKE_ATTEMPT_TIMEOUT_SEC=10s)
so a cold-starting process fails fast and retries instead of hanging until
the transport drops.
- Stop retrying ONLY when the process has DEFINITIVELY exited: new
_managed_process_has_exited() (checks EXITED status) replaces the previous
not-_managed_process_is_running() test, which false-negatived on a
just-spawned process that had not yet reported RUNNING and made the loop bail
to the outer rebuild path (relay then rejected the early re-attach with HTTP 400).
Adds a unit test that fails the first two handshakes with the process alive and
asserts the loop retries to success while starting the process exactly once.
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
* refactor(mcp): make MCP test reuse the shared Box session instead of a per-test session
Testing an MCP server (config-page "test" button) previously spun up a fresh
isolated mcp-test-<uuid> Box session every time: cold-start the container, run
the dependency bootstrap, probe, then tear the whole session down. That is slow
(tens of seconds) and, on an already-hosted server, wasteful — the server is
already running in the shared session.
Change the test to reuse the shared session / live process:
- _build_box_session_id: transient tests now use mcp-shared, the same Box
session as live servers, so a test reuses the running container (and, for an
existing server, its live managed process) instead of a cold per-test session.
- cleanup_session: a transient test no longer deletes the whole session (which
under the shared model would kill every other MCP server in the container). It
stops only its own process_id, exactly like a live server. Isolation is now at
the process level (distinct process_id per server/test), not the session level.
- test_mcp_server (persisted server): reuse the live connection with a real
list_tools refresh/probe; only fall back to a full start() when there is no
live connection to probe or the refresh fails, instead of an ERROR->start()
rebuild.
Trade-off: a failing test now shares the container with live servers rather than
a throwaway session. Accepted deliberately in favour of near-instant tests;
process-level isolation keeps a test from stopping another server's process.
* chore(deps): pin langbot-plugin 0.4.9 for the nsjail RLIMIT_AS node/npx MCP fix
---------
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
* fix(mcp): survive transient WS transport drops for Box stdio MCP servers
A Box-backed stdio MCP server (e.g. pab1it0/prometheus) would periodically
error on the frontend with Box managed process exited unexpectedly /
Failed after 4 attempts once the session had been alive for a while.
Root cause: the managed MCP process lives in the Box runtime and SURVIVES a
WebSocket transport drop, but _lifecycle_loop treated any monitor completion
as a fatal process death. It then ran the finally-block cleanup — which STOPS
the still-healthy managed process — and did a full 4-attempt exponential
backoff rebuild. Under an occasionally-stalled single-worker event loop the
mcp websocket client misses a ping/pong, the transport drops, and this
self-inflicted teardown loop is what the user sees.
Fixes:
- _lifecycle_loop: when the health monitor completes, re-check the real
managed-process state. If the process is still running, the transport
merely dropped: raise an internal _TransportReconnect signal instead of
Box managed process exited unexpectedly.
- _lifecycle_loop_with_retry: handle _TransportReconnect as a free, uncounted
reconnect (does not consume the fatal retry budget), so a long-lived session
survives arbitrarily many transient drops.
- finally-block: gate managed-process teardown on a _preserve_managed_process
flag so a transport-only reconnect closes just the WS, not the process.
- BoxStdioSessionRuntime.initialize: reuse an already-running managed process
instead of stopping+rebuilding it (which also re-ran the slow dependency
bootstrap); only (re)start when none is running. Adds
_managed_process_is_running() helper.
Pairs with langbot-plugin-sdk fix adding a server-driven WS heartbeat to the
managed-process relay, which prevents most drops in the first place.
* style(mcp): ruff format
* chore(deps): pin langbot-plugin 0.4.8 for the managed-process WS heartbeat fix
---------
Co-authored-by: dadachann <185672915+dadachann@users.noreply.github.com>
Hyu