An LLM call that returns neither text nor tool calls was recorded as a
success and the run completed with an empty reply, so users got nothing
and no error (seen with gemini-3.8-flash on LangBot Models, whose gateway
swallowed an upstream schema error into an empty 200 stream).
LocalAgent 0.2.12: when the first answer of a run is empty, fall back to
the next configured model; with no fallback left, fail the run with
runner.empty_response (retryable) so the pipeline error handling replies.
An empty normal stop after tools ran stays a valid completion.
Core: a successful LLM call with no output or zero reported usage keeps
status success but records details.warnings and a visible message on the
monitoring LLM call, and logs a warning.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
Strict chat templates served by vLLM/SGLang (Qwen3.5/3.6, Mistral, Gemma)
reject a system message that is not first, and some APIs (DeepSeek reasoner)
reject successive user/assistant messages. LocalAgent sent the knowledge-base
context as a mid-conversation system message and the event-data message as a
second user turn, so any KB-backed run on Qwen3.5+ via vLLM/SGLang failed with
"Invalid request parameters".
- Retrieved KB context is untrusted data and now uses the user role.
- Every model call normalizes roles at the boundary: all system messages fold
into one leading system message and consecutive user (or tool-call-free
assistant) messages merge. Tool call/result turns are never merged.
- Bump LocalAgent to 0.2.11.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(provider): decouple token counting from reasoning validation
Co-authored-by: multica-agent <github@multica.ai>
* fix(plugins): 修复 EssentialCommands !func 兼容性并清理 HelloPlugin 调试输出 (LANG-391) (#2967)
Co-authored-by: multica-agent <github@multica.ai>
* fix(plugin): wait for runtime recovery before workspace dispatch (#2968)
Co-authored-by: multica-agent <github@multica.ai>
* fix(runner): wait for installed plugins to initialize (LANG-393) (#2971)
Co-authored-by: multica-agent <github@multica.ai>
* fix(persistence): release SQLite writer during image refresh (#2973)
Co-authored-by: multica-agent <github@multica.ai>
* fix(telemetry): report cancellations as cancelled, keep route failure detail, tolerate already-terminal runs (#2974)
- Cancelled ingress / handed-off lanes / event deliveries close as `cancelled`
instead of `failed` (CancelledError and GeneratorExit).
- Every `runner_failed` event-route record carries the exception type and
message in `error`, including invalid Agent configuration.
- Closing the stream of a run another writer already finalized no longer
raises "Cannot transition terminal run"; the ledger status is kept and
reported, and the Runner's own error propagates.
- Platform API calls that return None without raising are `success`, not
`unknown`.
- Command lanes name their adapter and no longer claim the Pipeline's
configured Runner; only a Runner that actually ran is attributed.
LANG-395
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(agent): bound token counting and preserve text turns on failure
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: multica-agent <github@multica.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* refactor(agent): derive Runner input media from one ordered list
Every Runner input is now rendered once from the platform message chain
by agent/runner/input_media.py, replacing the contents/message-chain
cross matching, the descriptor merge from #2926 and the Box ref
reconciliation:
- contents place [attachment att-N: <type> <name>] at each media
position; quoted messages are wrapped in [quoted message from X] ...
[/quoted message] and their media marked source=quote.
- attachments describe the same media one-to-one as metadata only and
never carry inline content. Media added by user_message_alter stay
model-only.
- read_input_attachment(ref) serves bytes on demand (ExecutionAPI run
scope, available_apis.input_attachments): base64, http(s) URL,
adapter path, then adapter get_file_url for ID-only media; 10 MiB per
attachment, 50 MiB per run, no caching or payload logging.
- Event envelopes carry inline payloads Host-internally until
prepare_input_files registers them by ref; steering items drop refs
that would address the claimed run's own input.
- The media cache defaults to a 1024 MB cap (oldest removed first).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* docs(plugin): document input attachment refs and read_input_attachment
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(plugins): read input attachments by ref in official Runners
Dify, Coze, Tbox and ACP read attachment bytes with
ctx.read_input_attachment when the SDK and Host support it, upload each
media once under an att-N prefixed name, and skip unreadable ones;
older SDKs or Hosts keep the previous inline/URL/contents logic.
DeerFlow takes images from contents only on current Hosts, n8n forwards
attachment metadata, and LocalAgent exposes refs to the model.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* chore(deps): pin langbot-plugin 0.7.23 from PyPI
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* docs(plugin): note the SDK version for read_input_attachment
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(plugin): accept the SDK send_message timeout and route read_input_attachment
SDK 0.7.23 sends a client-side send_message timeout that the Runtime
strips; an older Runtime forwards it unchanged, so the Host now accepts
and ignores it instead of rejecting the call. The routing integration
test covers the new read_input_attachment operation.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
- Document workspace.language.get with exposure 'all' and risk 'read' so
plugins can declare it and Agents/Pipelines can select it as a workspace
tool. It requires workspace.view, takes no arguments and reads the
Workspace only from the trusted context. workspace.language.update stays
member/API-key only.
- Declare the operation in LocalAgent.
- Document it in zh/en/ja plugin Workspace API references and dev-docs.
- Cover grants, denial, invalid input, the unset default and the processor
tool path.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(provider): keep models resolvable during reloads and report quota errors clearly
Cloud catalog syncs triggered by any directory change called
load_models_from_db(), which emptied every tenant's runtime model cache
before reloading Workspaces one by one. Messages arriving in that window
could not resolve their configured model, and LocalAgent failed with
"No authorized model for local-agent".
- Stage the reload in a separate manager and swap it in atomically;
adopt unchanged providers instead of retiring and rebuilding them.
- Isolate per-Workspace load failures so one tenant cannot leave later
Workspaces unloaded.
- Lazily load a persisted model on a cache miss (another replica,
failed startup load, keys invalidated by a racing write).
- Classify New API quota errors (insufficient_user_quota,
pre_consume_token_quota_failed, HTTP 402, quota wording) as
quota_exceeded with a clear message in the Workspace language
(English by default) instead of "Permission denied"/"API error".
- Read upstream codes from the response body and exception chain,
and append safe machine codes to otherwise generic API errors.
- Report a missing API key explicitly instead of a raw OpenAIError.
- Keep HTTP status and body in the leased LiteLLM HTTP handler, which
previously turned every HTTP error into "Connection failed".
Refs LANG-318
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* fix(local-agent): accept bare model UUIDs and explain missing models
- parse_model_config accepts the bare model UUID shape the Host
selector already accepts, instead of treating it as "no model".
- Replace "No authorized model for local-agent" with an actionable
message pointing at the pipeline's model selection.
- Bump LocalAgent to 0.2.8.
Refs LANG-318
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* feat(api): open marketplace search and install to plugins and Agents
- Add plugin.marketplace.search, a read operation that proxies the configured
Space plugin search, honours plugin.enable_marketplace and returns only the
fields needed to install.
- Expose plugin.install.marketplace to plugins and processors as a dangerous
operation. Plugins and Agents cannot auto-confirm permissions or replace an
installed plugin; new installations stay disabled until a member confirms.
GitHub and local installs keep their current exposure.
- Expose task.get so callers can follow installation progress; non-member
principals receive the public projection, now including
pending_permission_confirmation.
- Declare the new operations in LocalAgent, document them in zh/en/ja and
localize the new errors.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
* test(api): include marketplace search and task status in processor read sweep
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: multica-agent <github@multica.ai>
Unified Workspace and execution APIs for LangBot (#2639).
- Versioned Application API shared by HTTP, MCP, the Workspace assistant
and plugins, with one authorization and error contract.
- Execution API facade for every plugin and Runner action.
- Plugin installation permissions: manifest declarations, member
confirmation, capability ceilings, compatibility mode, revocation and
audit. Typed Workspace API in the plugin SDK.
- Processor tool configuration: workspace operations, event and bot tools,
default tool sets for new Agents and Pipelines, and sensitive-tool
warnings.
- Plugin permission picker in the WebUI; LocalAgent declares every
plugin-callable permission.
- Documentation for plugin permissions and the Workspace API in English,
Chinese and Japanese.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
LocalAgent declared langbot-plugin>=0.7.18, and the published 0.2.2 artifact
pinned langbot-plugin==0.7.16. The plugin runtime rejects any SDK requirement
its own SDK version does not satisfy, so both directions were dead ends: 4.11.4
(SDK 0.7.19) could not install the published build, and a 0.7.18 floor would
have locked out 4.11.2/4.11.3 (SDK 0.7.16) from the next one.
Use a bounded range from the 4.11.2 baseline instead, matching every other
official Runner and KnowledgeEngine plugin. Verified on SDK 0.7.16 and 0.7.19:
327 tests pass on both, including the runtime dependency admission contract.
Co-authored-by: multica-agent <github@multica.ai>