mnajafian-nv · Oct 12, 2026 · Oct 8, 2026 · Oct 8, 2026 · Oct 8, 2026 · Oct 8, 2026
@@ -105,6 +105,12 @@ jobs:
           set -euo pipefail
           just --set no_uv true test-python
 
+      - name: Test OpenHands condensation with the SDK
+        if: matrix.platform == 'linux-amd64' && matrix.python-version == '3.12'
+        run: |
+          uv pip install "openhands-sdk==1.50.0" "openhands-tools==1.50.0"
+          uv run --no-sync pytest tests/adapters/test_openhands.py -k test_invoke_condenses_history_and_counts_summary_usage -q
+
   relay-0-9:
     name: Relay 0.9 (Python 3.12, linux-amd64)
     # The locks resolve Relay 0.10; this lane keeps the lowest supported minor

@@ -132,7 +132,7 @@ and additive extension maps because their support does not vary by adapter:
 | `schema_version` | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core |
 | `metadata.name`, `.description` | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core |
 | `harness.adapter_id`, `.resolution` | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core | Core |
-| `harness.settings` | Closed adapter schema | Closed adapter schema | Closed adapter schema | Closed adapter schema | Closed `timeout` schema | Closed schema | Closed command, port-range, and timeout schema | Closed empty schema | Closed schema with no properties | No settings declared | Closed local-extension and Relay-extension schema | Closed `permission_mode` schema | Closed schema | Closed `base_url`, `api_type`, transport-timeout, and `relay_streaming` schema |
+| `harness.settings` | Closed adapter schema | Closed adapter schema | Closed adapter schema | Closed adapter schema | Closed `timeout` schema | Closed schema | Closed command, port-range, and timeout schema | Closed `condenser` schema | Closed schema with no properties | No settings declared | Closed local-extension and Relay-extension schema | Closed `permission_mode` schema | Closed schema | Closed `base_url`, `api_type`, transport-timeout, and `relay_streaming` schema |
 | `workflow.target_id`, `.settings` | No | No | No | No | No | No | No | No | No | No | No | No | No | No |
 | `models.<role>.provider` | `anthropic` uses native auth; custom names require an Anthropic Messages-compatible `base_url` and `api_key_env` | `openai` uses native auth; custom names require a Responses-compatible `base_url` and `api_key_env` | Dynamic LangChain provider; custom OpenAI-compatible endpoints require `base_url` and `api_key_env` | Dynamic Hermes provider | Configured provider | Configured provider | OpenClaw provider; custom providers require an OpenAI Chat Completions-compatible `base_url` | Prefixes unqualified model names for LiteLLM | Cline provider | OpenCode provider | Pi catalog provider | `openai` | Kilo provider | Configured provider |
 | `models.<role>.model` | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes | Yes; passed to OpenCode | Yes; must exist in the Pi catalog | Yes | Yes | Yes |

@@ -74,6 +74,30 @@ directory. Each skill path must be a directory containing `SKILL.md`. The
 adapter disables ambient OpenHands user, public, project, and memory loading;
 only explicitly configured NeMo Fabric skills are loaded.
 
+The adapter configures the OpenHands SDK standard summarizing condenser,
+`default_condenser`, with the agent model. When the history sent to the model
+exceeds 80 events, or exceeds the model input token limit when OpenHands knows
+it, OpenHands replaces older events with a model-generated summary and keeps
+the first 4 events. When the model rejects a request for exceeding its context
+window, OpenHands condenses the history and continues instead of failing the
+invocation. Summarization calls count toward invocation usage, and each
+condensation step counts toward `runtime.max_turns`. NeMo Fabric does not
+expose condenser tuning parameters.
+
+To keep the full conversation history and make no summarization calls, set the
+`condenser` harness setting to `none`. The default value, `default`, enables
+the condenser, and any other value is rejected. The following example turns
+the condenser off:
+
+```python
+from nemo_fabric import HarnessConfig
+
+harness = HarnessConfig(
+    adapter_id="nvidia.fabric.openhands",
+    settings={"condenser": "none"},
+)
+```
+
 The initial adapter does not support Relay telemetry, native streaming,
 interactive confirmation, MCP authentication, or per-server MCP tool filters.
 

@@ -28,7 +28,14 @@
   "settings_schema": {
     "$schema": "https://json-schema.org/draft/2020-12/schema",
     "type": "object",
-    "properties": {},
+    "properties": {
+      "condenser": {
+        "type": "string",
+        "enum": ["default", "none"],
+        "default": "default",
+        "description": "OpenHands conversation history condenser. Use none to keep the full history."
+      }
+    },
     "additionalProperties": false
   },
   "requirements": {},

@@ -47,6 +47,7 @@ class OpenHandsApi(NamedTuple):
     TerminalTool: Any
     FileEditorTool: Any
     get_agent_final_response: Any
+    default_condenser: Any
 
 
 def _load_openhands_api() -> OpenHandsApi:
@@ -68,6 +69,7 @@ def _load_openhands_api() -> OpenHandsApi:
     from openhands.sdk import ConversationExecutionStatus
     from openhands.sdk import LLM
     from openhands.sdk import Tool
+    from openhands.sdk.context.condenser import default_condenser
     from openhands.sdk.conversation import get_agent_final_response
     from openhands.sdk.conversation.impl.local_conversation import LocalConversation
     from openhands.sdk.mcp import MCPServer
@@ -88,6 +90,7 @@ def _load_openhands_api() -> OpenHandsApi:
         TerminalTool=TerminalTool,
         FileEditorTool=FileEditorTool,
         get_agent_final_response=get_agent_final_response,
+        default_condenser=default_condenser,
     )
 
 
@@ -128,6 +131,23 @@ def _llm(api: OpenHandsApi, model: contract.AgentModelConfig) -> Any:
     return api.LLM(**kwargs)
 
 
+def _condenser_enabled(config: contract.AgentConfig) -> bool:
+    """Return whether to condense history, rejecting values the descriptor forbids.
+
+    Planning already validates the setting, but a host that skips planning must
+    not turn an unknown value into model-billed summarization calls.
+    """
+
+    settings = config.harness.settings if config.harness is not None else {}
+    condenser = settings.get("condenser", "default")
+    if condenser not in ("default", "none"):
+        raise lifecycle.LifecycleError(
+            "openhands_condenser_invalid",
+            "The OpenHands condenser setting must be 'default' or 'none'.",
+        )
+    return condenser == "default"
+
+
 def _workspace(context: contract.RuntimeContext, base_dir: str) -> Path:
     path = Path(context.environment.workspace or base_dir)
     if not path.is_absolute():
@@ -343,6 +363,10 @@ async def start(self, payload: dict[str, Any]) -> None:
             "mcp_config": _mcp_servers(api, config),
             "agent_context": agent_context,
         }
+        if _condenser_enabled(config):
+            # Share the agent LLM instead of a copy so summarization calls
+            # accrue to the metrics that invoke() reports as usage.
+            agent_kwargs["condenser"] = api.default_condenser(llm)
         if instruction is not None and instruction.mode == "replace":
             agent_kwargs["system_prompt"] = instruction.content
         agent = api.Agent(**agent_kwargs)

@@ -68,6 +68,30 @@ Skill paths resolve from the configuration base directory and must contain a
 `SKILL.md` file. The adapter disables ambient OpenHands user, public, project,
 and memory loading, so only explicitly configured NeMo Fabric skills load.
 
+The adapter configures the OpenHands SDK standard summarizing condenser,
+`default_condenser`, with the agent model. When the history sent to the model
+exceeds 80 events, or exceeds the model input token limit when OpenHands knows
+it, OpenHands replaces older events with a model-generated summary and keeps
+the first 4 events. When the model rejects a request for exceeding its context
+window, OpenHands condenses the history and continues instead of failing the
+invocation. Summarization calls count toward invocation usage, and each
+condensation step counts toward `runtime.max_turns`. NeMo Fabric does not
+expose condenser tuning parameters.
+
+To keep the full conversation history and make no summarization calls, set the
+`condenser` harness setting to `none`. The default value, `default`, enables
+the condenser, and any other value is rejected. The following example turns
+the condenser off:
+
+```python
+from nemo_fabric import HarnessConfig
+
+harness = HarnessConfig(
+    adapter_id="nvidia.fabric.openhands",
+    settings={"condenser": "none"},
+)
+```
+
 The initial integration does not support Relay telemetry, native streaming,
 interactive confirmation, MCP authentication, or per-server MCP tool filters.
 

@@ -1664,7 +1664,17 @@
         "settings_schema": {
           "$schema": "https://json-schema.org/draft/2020-12/schema",
           "additionalProperties": false,
-          "properties": {},
+          "properties": {
+            "condenser": {
+              "default": "default",
+              "description": "OpenHands conversation history condenser. Use none to keep the full history.",
+              "enum": [
+                "default",
+                "none"
+              ],
+              "type": "string"
+            }
+          },
           "type": "object"
         }
       },
@@ -1673,7 +1683,7 @@
         "name": "nemo-fabric-adapters-openhands",
         "version": "0.5.0"
       },
-      "sha256": "151335fc5bbc1e1689112b0629aa0e517e4de9b4a5a3ba52cfa4a8ec7001ae7d",
+      "sha256": "062ef09ba5e82dfdf0b4374632a8b39b7ff975160b0c6f62e68d2f4ac6f0b59e",
       "source": "adapters/python/openhands/openhands.fabric-adapter.json"
     },
     "nvidia.fabric.pi": {

@@ -124,6 +124,7 @@ async def arun():
     conversation.arun = AsyncMock(side_effect=arun)
     conversation.close = MagicMock()
     conversation_factory = MagicMock(return_value=conversation)
+    default_condenser = MagicMock()
 
     api = adapter.OpenHandsApi(
         Agent=agent_factory,
@@ -138,6 +139,7 @@ async def arun():
         TerminalTool=SimpleNamespace(name="TerminalTool"),
         FileEditorTool=SimpleNamespace(name="FileEditorTool"),
         get_agent_final_response=lambda events: events[-1] if events else "",
+        default_condenser=default_condenser,
     )
     monkeypatch.setattr(adapter, "_load_openhands_api", lambda: api)
     calls.update(
@@ -146,6 +148,7 @@ async def arun():
             "agent_factory": agent_factory,
             "conversation": conversation,
             "conversation_factory": conversation_factory,
+            "default_condenser": default_condenser,
             "llm_factory": llm_factory,
             "mcp_server_factory": mcp_server_factory,
             "skill_type": skill_type,
@@ -199,6 +202,18 @@ def test_descriptor_declares_only_initial_openhands_surface():
         "updates": False,
         "cancellation": False,
     }
+    assert descriptor["settings_schema"]["properties"] == {
+        "condenser": {
+            "type": "string",
+            "enum": ["default", "none"],
+            "default": "default",
+            "description": (
+                "OpenHands conversation history condenser. Use none to keep "
+                "the full history."
+            ),
+        }
+    }
+    assert descriptor["settings_schema"]["additionalProperties"] is False
     assert "telemetry" not in descriptor
 
 
@@ -281,6 +296,153 @@ async def test_start_maps_normalized_configuration(
     assert profile_store.is_dir()
 
 
+async def test_start_condenses_history_with_the_agent_llm(
+    openhands_payload: dict,
+    mock_openhands: dict,
+    monkeypatch: pytest.MonkeyPatch,
+):
+    monkeypatch.setenv("TEST_OPENHANDS_API_KEY", "test-key")
+    await adapter.OpenHandsRuntime().start(_start_payload(openhands_payload))
+
+    llm = mock_openhands["llm_factory"].return_value
+    default_condenser = mock_openhands["default_condenser"]
+    default_condenser.assert_called_once_with(llm)
+    agent_kwargs = mock_openhands["agent_factory"].call_args.kwargs
+    assert agent_kwargs["condenser"] is default_condenser.return_value
+
+
+async def test_start_without_condenser_when_disabled(
+    openhands_payload: dict,
+    mock_openhands: dict,
+    monkeypatch: pytest.MonkeyPatch,
+):
+    monkeypatch.setenv("TEST_OPENHANDS_API_KEY", "test-key")
+    openhands_payload["config"]["harness"] = {"settings": {"condenser": "none"}}
+    await adapter.OpenHandsRuntime().start(_start_payload(openhands_payload))
+
+    mock_openhands["default_condenser"].assert_not_called()
+    agent_kwargs = mock_openhands["agent_factory"].call_args.kwargs
+    assert agent_kwargs.get("condenser") is None
+
+
+@pytest.mark.parametrize(
+    "condenser", ["off", "None", None], ids=["off", "string-None", "null"]
+)
+async def test_start_rejects_unknown_condenser_setting(
+    openhands_payload: dict,
+    mock_openhands: dict,
+    monkeypatch: pytest.MonkeyPatch,
+    condenser: str | None,
+):
+    monkeypatch.setenv("TEST_OPENHANDS_API_KEY", "test-key")
+    openhands_payload["config"]["harness"] = {"settings": {"condenser": condenser}}
+
+    with pytest.raises(adapter.lifecycle.LifecycleError) as caught:
+        await adapter.OpenHandsRuntime().start(_start_payload(openhands_payload))
+
+    assert caught.value.code == "openhands_condenser_invalid"
+    mock_openhands["default_condenser"].assert_not_called()
+    mock_openhands["agent_factory"].assert_not_called()
+
+
+# Event limit: the system prompt, the user message, and 40 think action and
+# observation pairs exceed the 80-event limit, so the condenser runs before the
+# 41st agent call. Context window error: the rejected 11th call triggers the
+# condensation, and the 12th call finishes.
+@pytest.mark.parametrize(
+    ("context_error_at_call", "expected_agent_calls"),
+    [(None, 41), (11, 12)],
+    ids=["event-limit", "context-window-error"],
+)
+async def test_invoke_condenses_history_and_counts_summary_usage(
+    openhands_payload: dict,
+    monkeypatch: pytest.MonkeyPatch,
+    context_error_at_call: int | None,
+    expected_agent_calls: int,
+):
+    pytest.importorskip("openhands.sdk")
+    pytest.importorskip("openhands.tools")
+    from litellm.exceptions import ContextWindowExceededError
+    from litellm.types.utils import ModelResponse
+    from openhands.sdk.event import Condensation
+    from openhands.sdk.llm import llm as openhands_llm
+
+    calls = {"agent": 0, "summary": 0}
+
+    def respond(**kwargs) -> ModelResponse:
+        # Agent steps send tools; the condenser's summary request does not.
+        if kwargs.get("tools"):
+            calls["agent"] += 1
+            if calls["agent"] == context_error_at_call:
+                raise ContextWindowExceededError(
+                    message="context length exceeded",
+                    model="test-model",
+                    llm_provider="openai",
+                )
+            tool, arguments = (
+                # Distinct thoughts keep the SDK stuck detector quiet.
+                ("think", {"thought": f"Step {calls['agent']}."})
+                if calls["summary"] == 0
+                else ("finish", {"message": "done"})
+            )
+            message = {
+                "role": "assistant",
+                "content": "",
+                "tool_calls": [
+                    {
+                        "id": f"call-{calls['agent']}",
+                        "type": "function",
+                        "function": {"name": tool, "arguments": json.dumps(arguments)},
+                    }
+                ],
+            }
+            prompt_tokens, completion_tokens = 10, 1
+        else:
+            calls["summary"] += 1
+            message = {"role": "assistant", "content": "Summary of earlier steps."}
+            prompt_tokens, completion_tokens = 1000, 100
+        return ModelResponse(
+            model="test-model",
+            choices=[{"index": 0, "finish_reason": "stop", "message": message}],
+            usage={
+                "prompt_tokens": prompt_tokens,
+                "completion_tokens": completion_tokens,
+                "total_tokens": prompt_tokens + completion_tokens,
+            },
+        )
+
+    monkeypatch.setattr(
+        openhands_llm, "litellm_completion", MagicMock(side_effect=respond)
+    )
+    monkeypatch.setattr(
+        openhands_llm, "litellm_acompletion", AsyncMock(side_effect=respond)
+    )
+    monkeypatch.setenv("TEST_OPENHANDS_API_KEY", "test-key")
+    config = openhands_payload["config"]
+    config["tools"] = {"enabled": [], "blocked": []}
+    config["runtime"] = {"max_turns": 100}
+    del config["skills"], config["mcp"]
+    runtime = adapter.OpenHandsRuntime()
+    await runtime.start(_start_payload(openhands_payload))
+    try:
+        result = await runtime.invoke(
+            *_invocation(openhands_payload, "Think until the history condenses.")
+        )
+        events = list(runtime._conversation.state.events)
+    finally:
+        await runtime.stop()
+
+    assert result.status is AgentRunStatus.SUCCEEDED
+    assert result.output == {"response": "done"}
+    assert calls["summary"] == 1
+    assert calls["agent"] == expected_agent_calls
+    assert sum(isinstance(event, Condensation) for event in events) == 1
+    # A rejected request reports no usage; every other call counts once.
+    answered_calls = calls["agent"] - (context_error_at_call is not None)
+    assert result.usage.input_tokens == 10 * answered_calls + 1000
+    assert result.usage.output_tokens == answered_calls + 100
+
+
 async def test_replace_instruction_uses_exact_system_prompt(
     openhands_payload: dict,
     mock_openhands: dict,