Plugin Daemon Model Parameter Handling#

When an Agent node invokes a model, it must send a valid completion_params (model parameter) payload to the plugin daemon. A structural divergence between how the workflow Agent node and the standalone Agent app build this payload was the root cause of InvokeError: [models] Error: 'required' failures reported in issue #39274.

The Bug: Empty completion_params in Workflow Agent Nodes#

The workflow frontend's model-selector widget only persists provider, model, and mode — it does not store completion_params. When the plugin SDK's AgentModelConfig receives a value without completion_params, it defaults the field to {}. 1

Agent Strategy plugins (cot_agent, function_calling) use completion_params to construct the LLMModelConfig passed back to the plugin daemon for the actual LLM call. Some model providers (e.g. Tongyi Qwen, any provider with parameter_rules that declare required fields) call _validate_and_filter_model_parameters, which raises KeyError('required') when the parameters dict is empty. 1

The standalone Agent app does not hit this path because it populates model settings from agent_soul.model.model_settings, which are then sanitized and forwarded through _agent_model_settings(). Known parameters (temperature, top_p, presence_penalty, frequency_penalty, max_tokens, stop_sequences) are passed directly, and any additional plugin-specific parameters are forwarded via extra_body. 2

The Fix: Populate Defaults from Model Schema#

PR #39590 (merged 2026-07-28) fixed this in AgentRuntimeSupport.build_parameters(). When the model-selector value lacks completion_params, it now populates it from the selected model's schema defaults:

• If a model_schema (AIModelEntity) is available: calls _extract_default_completion_params(model_schema) which iterates model_schema.parameter_rules and collects every rule's default value where default is not None. 3
• If no schema is available: falls back to {}. 4

A prior attempt (PR #39373) tried renaming completion_params to model_parameters in build_parameters(), but this was the wrong field name — AgentModelConfig expects completion_params — so the parameters were dropped instead of forwarded. 5

AgentSoulModelSettings: Persistence of Plugin Parameters#

The AgentSoulModelSettings model (defined in api/models/agent_config_entities.py) validates and stores model parameters for Agent Soul configurations. It now uses extra="allow" (via AgentFlexibleConfig inheritance) instead of the previous extra="ignore", meaning plugin-declared parameters (such as Qwen's enable_thinking or Tongyi-specific settings) are preserved during validation and persist through save/publish round trips, rather than being silently dropped.

This change ensures that parameters declared by model plugins via parameter_rules survive the configuration lifecycle. Combined with the _agent_model_settings() forwarding mechanism, plugin-specific parameters can now flow from the UI through persistence into runtime execution.

Parameter Flow by Path#

For Agent V2 and the standalone app, the model_settings dict passes through _agent_model_settings() in api/clients/agent_backend/request_builder.py, which explicitly forwards temperature, top_p, presence_penalty, frequency_penalty, max_tokens, and stop_sequences. Additionally, any unknown parameters (such as plugin-specific ones like Qwen's enable_thinking) are preserved and forwarded via the extra_body dictionary, ensuring plugin-declared model parameters reach the backend. 8

Legacy V1 Agent nodes use AgentNodeData and the build_parameters() path. The MODEL_SELECTOR parameter type is the trigger point where completion_params extraction occurs. 3

Key Source Files#

Debugging Tips#

• The 'required' string comes from the plugin daemon's _validate_and_filter_model_parameters, which performs JSON schema validation against the model provider's parameter_rules. Check docker compose logs plugin_daemon --tail 100 for the full stack trace. 9
• The error manifests in Chatflow/workflow context but not in standalone Agent app runs, because the standalone path populates model settings before they reach the plugin daemon. 9
• After applying the fix (v1.16.x+), agents running inside Chatflow/workflow nodes receive completion_params populated from the model's declared defaults, satisfying provider-level required field validation.
• If upgrading is not immediately possible, manually re-saving the Agent node's model selection in the Chatflow editor can sometimes force a fresh configuration snapshot that avoids stale empty completion_params.