Add SDK-backed OpenHands SFT converter (#241)

* Add SDK-backed OpenHands SFT converter

* Normalize SDK SFT records for HF chat templates

* Keep SDK dependencies optional for CI

* Bump schema version for dataset metadata

* Run CI on Python 3.12 for SDK support

* Address SDK converter review feedback

* Preserve webshop custom click labels

---------

Co-authored-by: Graham Neubig <gneubig@users.noreply.github.com>

Author: neubig

.github/workflows/check_api_docstrings.yml

@@ -19,7 +19,7 @@ jobs:
     - name: Set up Python
       uses: actions/setup-python@v5
       with:
-        python-version: '3.11'
+        python-version: '3.12'
 
     - name: Install ruff
       run: |

.github/workflows/pre-commit.yml

@@ -17,7 +17,7 @@ jobs:
       - name: Set up Python
         uses: actions/setup-python@v4
         with:
-          python-version: '3.10'
+          python-version: '3.12'
           cache: 'pip'
 
       - name: Install dependencies

.github/workflows/pytest.yml

@@ -11,7 +11,7 @@ jobs:
     runs-on: ubuntu-latest
     strategy:
       matrix:
-        python-version: [3.11]
+        python-version: [3.12]
 
     steps:
     - uses: actions/checkout@v4

agents/openhands_sdk/README.md

@@ -0,0 +1,24 @@
+# OpenHands SDK SFT Converter
+
+This converter builds SFT records through OpenHands SDK primitives instead of
+manually formatting OpenAI chat messages.
+
+For each standardized trajectory it:
+
+1. Loads dataset `metadata.json`.
+2. Registers every metadata `custom_tools` entry as an SDK custom tool.
+3. Creates an SDK `Conversation`, letting the SDK construct the system prompt
+   and tool map.
+4. Appends converted standardized actions and observations to the SDK event log.
+5. Uses SDK event/message formatting helpers to emit `messages` and `tools`.
+
+Regenerate one dataset with:
+
+```bash
+export MY_DATASET=agenttuning_alfworld
+export PYTHONPATH=`pwd`:$PYTHONPATH
+python scripts/json_to_jsonl.py < datasets/$MY_DATASET/sample_std.json \
+  | python agents/openhands_sdk/std_to_sft.py \
+  | python scripts/jsonl_to_json.py \
+  > datasets/$MY_DATASET/sample_sft/openhands_sdk.json
+```

agents/openhands_sdk/__init__.py

@@ -0,0 +1 @@
+"""OpenHands SDK SFT conversion package."""