scan_skill_directory is disabled for Cloud-managed sandboxes, but the
register_skill tool was still advertised to the model whenever a backend was
available, so every call failed closed without a usable explanation. Advertise
register_skill only when the runtime may scan sandbox directories, and keep
activate available in both modes.
The workspace file tools (read/write/edit/glob/grep) ran a script through
`python`, which a sandbox built from a host rootfs may not provide: those hosts
ship `python3` only, and a read-only /usr prevents creating a `python` shim
inside the jail. Every file tool then failed with `/bin/sh: 1: python: not
found`. Resolve the interpreter with `command -v python3 || command -v python`,
matching the interpreter the Box service already uses for its own scripts.
* 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 <[email protected]>
* feat(mcp): surface OAuth-required server tests
* fix(mcp): show connection failure details in status cards
---------
Co-authored-by: RockChinQ <[email protected]>
The Ollama requester declared litellm_provider: ollama, which routes
every request through litellm's legacy /api/generate-based
OllamaConfig. That config's get_supported_openai_params() does not
include "tools"/"tool_choice" at all, so an Ollama-hosted model in a
local-agent pipeline could never receive a structured tool definition
or return a structured tool_calls response - it could only try to
express a tool call as free text (typically inside its own <think>
reasoning), which LangBot then has no way to execute.
litellm's "ollama_chat" provider targets Ollama's modern /api/chat
endpoint instead, which correctly forwards tools/tool_choice and
correctly surfaces the model's native message.tool_calls field.
Verified against a real local Ollama 0.33.2 instance with the exact
system prompt, RAG-augmented user message, and tool set a live
pipeline sends.
Two follow-on fixes needed because the Ollama requester definition is
shared by LLM and text-embedding models:
- get_reasoning_capabilities: match family in ('ollama', 'ollama_chat')
so the reasoning-level UI still works for this provider.
- scan_models: retry {base_url}/v1/models on a 404 from {base_url}/models,
since Ollama's base_url is a bare host (must not include /v1 - that
would break OllamaChatConfig.get_complete_url, which appends /api/chat
to it directly), unlike most other OpenAI-compatible providers whose
base_url already ends in /v1.
- invoke_embedding: litellm's embedding routing has no "ollama_chat"
case, only "ollama". Build the embedding model name with an explicit
custom_llm_provider="ollama" override when the requester is configured
for ollama_chat, so embedding models (e.g. bge-m3) keep working.
Co-authored-by: zx90316 <[email protected]>
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.
RockChinQ