React Agent

Configurations (Graphs)

graph - Basic graph configuration without CCRS-specific nodes.

graph_ccrs - Graph configuration with CCRS.

Run

Can be run either via the commandline tool or the notebook.

Install dependencies

python -m pip install -r requirements.txt

For opportunistic CCRS, see the React adapter notes in react_agent/ccrs/README.md.

Notebook

Notebook

Use the notebook as the maintained easy-to-run place for predefined agent run configurations and implementation variants, including the baseline graph, opportunistic CCRS graph, and future contingency CCRS variants.

Commandline tool

python main.py --graph-name graph_ccrs --agent-name "CCRSAgent" --log-level "DEBUG"

Expose the optional LLM self-escalation tool for contingency CCRS:

python main.py --graph-name graph_ccrs --enable-contingency-escalation-tool --agent-name "CCRSAgent" --log-level "DEBUG"

Enable optional Java contingency providers from the CLI:

python main.py --graph-name graph_ccrs --enable-contingency-escalation-tool --enable-contingency-llm-prediction --sync-contingency-llm-model --agent-name "CCRSAgent" --log-level "DEBUG"

For A2A consultation as well:

python main.py --graph-name graph_ccrs --enable-contingency-escalation-tool --enable-contingency-llm-prediction --enable-contingency-a2a-consultation --sync-contingency-llm-model --agent-name "CCRSAgent" --log-level "DEBUG"

Options

--agent-name            (default: "React")

--log-level             Logging level (default: "INFO")

--run-mode              sync or async (default: "sync")

--graph-name            (default: "graph")

--recursion-limit       Override recursion limit

--query                 (default: Maze Prompt)

--llm-message-window-max-messages
                        Maximum recent non-preserved messages sent through the LLM message history window

--llm-message-window-max-tokens
                        Maximum approximate tokens for non-preserved messages sent through the LLM message history window

--enable-contingency-escalation-tool
                        Expose the opt-in escalate_to_contingency_ccrs tool when using graph_ccrs

--contingency-http-error-threshold
                        Consecutive HTTP status >= 400 tool responses before default contingency escalation

--enable-contingency-llm-prediction
                        Enable the optional Java contingency LLM prediction capability when using graph_ccrs

--enable-contingency-a2a-consultation
                        Enable the optional Java contingency A2A consultation capability when using graph_ccrs

--contingency-ccrs-modules
                        Comma- or space-separated Java CCRS modules for contingency evaluation,
                        for example: ccrs-core,ccrs-langchain4j,ccrs-a2a

--discover-contingency-strategy-providers
                        Discover Java contingency strategy providers with ServiceLoader when using graph_ccrs

--sync-contingency-llm-model
                        Set OPENAI_MODEL from the Python agent llm_model before constructing Java contingency providers

LLM settings such as model, temperature, reasoning effort, message-window limits, and LangChain project are read from .env through settings.py. The message window keeps the first user query and trims only the message history passed through the prompt's MessagesPlaceholder; system prompts and CCRS prompt text remain outside that trimming scope. Configure LLM_MESSAGE_WINDOW_MAX_MESSAGES for message-count trimming and LLM_MESSAGE_WINDOW_MAX_TOKENS for approximate token trimming.

References

• Google Gemini LangGraph Example

"LangGraph provides a convenient helper, add_messages, for updating message lists in the state. It functions as a reducer, meaning it takes the current list and new messages, then returns a combined list. It smartly handles updates by message ID and defaults to an "append-only" behavior for new, unique messages."

• LangGraph Documentation