langchain driver for OW. bug fixes along the way

align_system/configs/driver/chat_model/ollama_llama31.yaml

@@ -0,0 +1,4 @@
+_target_: langchain_ollama.ChatOllama
+model: llama3.1:latest
+temperature: 0.0
+num_ctx: 16384

align_system/configs/driver/chat_model/ollama_qwen25_32b.yaml

@@ -0,0 +1,4 @@
+_target_: langchain_ollama.ChatOllama
+model: qwen2.5:32b
+temperature: 0.0
+num_ctx: 16384

align_system/configs/driver/chat_model/ollama_qwen25_32b_itm.yaml

@@ -0,0 +1,7 @@
+# qwen2.5:32b served by the ollama instance on the itm server (4x RTX
+# A6000); reachable directly on the local network
+_target_: langchain_ollama.ChatOllama
+model: qwen2.5:32b
+base_url: http://itm:11434
+temperature: 0.0
+num_ctx: 16384

align_system/configs/driver/chat_model/ollama_qwen25_7b.yaml

@@ -0,0 +1,4 @@
+_target_: langchain_ollama.ChatOllama
+model: qwen2.5:7b
+temperature: 0.0
+num_ctx: 16384

align_system/configs/driver/chat_model/ollama_qwen3_32b.yaml

@@ -0,0 +1,13 @@
+_target_: langchain_ollama.ChatOllama
+model: qwen3:32b
+temperature: 0.0
+# qwen3's native context window; the OW scene observations are large,
+# and overflowing num_ctx makes ollama silently truncate the oldest
+# context (including the system prompt's instructions)
+num_ctx: 40960
+# qwen3 is a thinking model.  With reasoning false (think
+# suppression) qwen3:32b was observed returning completely empty
+# responses (no content, no tool calls); with true, the think phase
+# runs but langchain-ollama routes it to a separate channel, keeping
+# message content clean
+reasoning: true

align_system/configs/driver/chat_model/vllm_qwen25_15b.yaml

@@ -0,0 +1,9 @@
+# HuggingFace model served by vLLM's OpenAI-compatible endpoint (start
+# the server yourself; see vllm_qwen25_7b for details):
+#   vllm serve Qwen/Qwen2.5-1.5B-Instruct --enable-auto-tool-choice \
+#     --tool-call-parser hermes --max-model-len 8192 --port 8000
+_target_: langchain_openai.ChatOpenAI
+model: Qwen/Qwen2.5-1.5B-Instruct
+base_url: http://localhost:8000/v1
+api_key: EMPTY
+temperature: 0.0

align_system/configs/driver/chat_model/vllm_qwen25_3b.yaml

@@ -0,0 +1,13 @@
+# HuggingFace model served by vLLM's OpenAI-compatible endpoint (start
+# the server yourself; see vllm_qwen25_7b for details).  Qwen2.5-3B is
+# the smallest of the Qwen2.5 instruct models observed reliably
+# driving the agent loop with structured tool calls (1.5B tends to
+# emit dangling <tool_call> tags instead); at max-model-len 4096 it
+# fits an 8GB GPU:
+#   vllm serve Qwen/Qwen2.5-3B-Instruct --enable-auto-tool-choice \
+#     --tool-call-parser hermes --max-model-len 4096 --port 8000
+_target_: langchain_openai.ChatOpenAI
+model: Qwen/Qwen2.5-3B-Instruct
+base_url: http://localhost:8000/v1
+api_key: EMPTY
+temperature: 0.0

align_system/configs/driver/chat_model/vllm_qwen25_7b.yaml

@@ -0,0 +1,13 @@
+# HuggingFace model served by vLLM's OpenAI-compatible endpoint.
+# Serving is left to the user (the align system doesn't manage the
+# server process); start it before running, with tool calling enabled:
+#   vllm serve Qwen/Qwen2.5-7B-Instruct --enable-auto-tool-choice \
+#     --tool-call-parser hermes --port 8000
+# See the docstring of align_system.drivers.itm_open_world_langchain
+# for why the agent goes through the server rather than
+# align_system.algorithms.vllm_inference_engine.
+_target_: langchain_openai.ChatOpenAI
+model: Qwen/Qwen2.5-7B-Instruct
+base_url: http://localhost:8000/v1
+api_key: EMPTY
+temperature: 0.0

align_system/configs/driver/itm_phase2_ow_langchain.yaml

@@ -0,0 +1,36 @@
+defaults:
+  # Chat model for the agent -- must support tool calling.  Select a
+  # different backend with e.g.
+  #   - override /driver/chat_model: vllm_qwen25_7b
+  # in an experiment config (see configs/driver/chat_model/ for the
+  # available options; add new yamls there for other providers such as
+  # langchain_openai.ChatOpenAI or langchain_anthropic.ChatAnthropic).
+  # Alternatively set `model` to an init_chat_model string (e.g.
+  # "openai:gpt-4o"), which takes precedence over `chat_model`.
+  #
+  # NOTE: a local transformers pipeline
+  # (langchain_huggingface.HuggingFacePipeline) does NOT support tool
+  # calling and will not work with this driver; to run HuggingFace
+  # models locally, use the vllm_* chat_model configs instead, which
+  # talk to a separately started `vllm serve` OpenAI-compatible
+  # endpoint (see the docstring of
+  # align_system.drivers.itm_open_world_langchain for why)
+  - chat_model: ollama_llama31
+  - _self_
+
+_target_: align_system.drivers.itm_open_world_langchain.ITMOpenWorldLangChainDriver
+
+model: null
+
+apply_action_filtering: true
+sort_available_actions: false
+
+# Hard cap on environment actions per scenario (safety net against
+# agent loops); when hit, the driver ends the scene itself
+max_actions_per_scenario: 100
+# If the agent makes this many LLM calls without taking an action, the
+# driver takes the first available action on its behalf so a live
+# session can't stall
+max_llm_calls_between_actions: 8
+# Rolling window of conversation messages kept in the agent's context
+max_messages_in_context: 40

align_system/configs/experiment/phase2_feb_openworld/phase2_langchain_agent_ow_live.yaml

@@ -0,0 +1,24 @@
+# @package _global_
+defaults:
+  # The LangChain agent drives the scenario itself; no ADM is used
+  - override /adm: null
+  - override /interface: ta3
+  - override /driver: itm_phase2_ow_langchain
+  # llama3.1-8b tends to narrate instead of emitting tool calls on the
+  # larger open-world scenes; qwen2.5:32b is a much more reliable
+  # tool-caller.  If you change the model, update interface.username
+  # to match (keep the testrun- prefix for test runs).
+  - override /driver/chat_model: ollama_qwen25_32b
+
+interface:
+#  api_endpoint: "https://darpaitm.caci.com"
+  api_endpoint: 'http://127.0.0.1:8081'
+  session_type: eval
+  training_session: null
+  username: "testrun-ALIGN-ADM-Ph2-LangChainAgent-Qwen2.5-32B-Instruct"
+  domain: "owtriage"
+  adm_profile: FEB_OPENWORLD3
+
+force_determinism: true
+align_to_target: true
+save_last_unstructured_state_per_scenario: true

align_system/configs/experiment/phase2_feb_openworld/phase2_langchain_agent_ow_live_hf.yaml

@@ -0,0 +1,27 @@
+# @package _global_
+# Live eval with the LangChain agent driver running a HuggingFace
+# model served by vLLM's OpenAI-compatible endpoint.  Start the server
+# first, e.g.:
+#   vllm serve Qwen/Qwen2.5-7B-Instruct --enable-auto-tool-choice \
+#     --tool-call-parser hermes --port 8000
+# If you serve a different model, override driver.chat_model.model and
+# update interface.username to match (keep the testrun- prefix for
+# test runs).
+defaults:
+  # The LangChain agent drives the scenario itself; no ADM is used
+  - override /adm: null
+  - override /interface: ta3
+  - override /driver: itm_phase2_ow_langchain
+  - override /driver/chat_model: vllm_qwen25_7b
+
+interface:
+  api_endpoint: "https://darpaitm.caci.com"
+  session_type: eval
+  training_session: null
+  username: "testrun-ALIGN-ADM-Ph2-LangChainAgent-HF-Qwen2.5-7B-Instruct"
+  domain: "owtriage"
+  adm_profile: FEB_OPENWORLD3
+
+force_determinism: true
+align_to_target: true
+save_last_unstructured_state_per_scenario: true

align_system/configs/experiment/phase2_feb_openworld/phase2_langchain_agent_ow_replay.yaml

@@ -0,0 +1,15 @@
+# @package _global_
+# Local smoke test for the LangChain agent driver: replays recorded
+# scenario states from a file instead of hitting a live TA3 server, so
+# the agent's observe -> decide -> act loop can be exercised offline
+defaults:
+  # The LangChain agent drives the scenario itself; no ADM is used
+  - override /adm: null
+  - override /interface: input_output_file
+  - override /driver: itm_phase2_ow_langchain
+
+interface:
+  input_output_filepath: 'example_data/input_output_files/openworld_smoke_input_output.json'
+  state_hydration_domain: p2triage
+
+align_to_target: false

align_system/configs/experiment/phase2_feb_openworld/phase2_langchain_agent_ow_replay_hf.yaml

@@ -0,0 +1,16 @@
+# @package _global_
+# Offline smoke test of the LangChain agent driver with a HuggingFace
+# model served by vLLM's OpenAI-compatible endpoint.  Start the server
+# first (see the vllm_* chat_model configs for the exact command)
+defaults:
+  # The LangChain agent drives the scenario itself; no ADM is used
+  - override /adm: null
+  - override /interface: input_output_file
+  - override /driver: itm_phase2_ow_langchain
+  - override /driver/chat_model: vllm_qwen25_3b
+
+interface:
+  input_output_filepath: 'example_data/input_output_files/openworld_smoke_input_output.json'
+  state_hydration_domain: p2triage
+
+align_to_target: false