Skip to content

README.md

Latest commit

 

History

History
184 lines (148 loc) · 7.4 KB

README.md

File metadata and controls

184 lines (148 loc) · 7.4 KB
Raw

NVIDIA NeMo Fabric mini-SWE-agent Adapter

Use the nvidia.fabric.mini-swe-agent adapter to run mini-SWE-agent with NVIDIA NeMo Fabric.

Install

Choose an installation option based on the components required in the target environment:

Installation Runtime Adapter Harness Relay Python
pip install "nemo-fabric[mini-swe-agent]" Yes Yes Yes No
pip install "nemo-fabric[mini-swe-agent,relay]" Yes Yes Yes Yes
pip install "nemo-fabric-adapters-mini-swe-agent[harness]" No Yes Yes No
pip install "nemo-fabric-adapters-mini-swe-agent[relay]" No Yes No Yes
pip install "nemo-fabric-adapters-mini-swe-agent[full]" No Yes Yes Yes
pip install nemo-fabric-adapters-mini-swe-agent No Yes No No

The harness and full extras install the latest compatible mini-SWE-agent 2.x release.

Configuration

The adapter supports models, models.base_url, models.temperature, models.top_p, models.max_tokens, replacement instructions.system, runtime.max_turns, and environment.workspace. It passes the normalized sampling fields through mini-SWE-agent's LiteLLM model arguments and rejects append system instructions. runtime.timeout_seconds sets the NVIDIA NeMo Fabric invocation deadline. Use harness.settings.timeout to set the maximum duration of one command; the default is 30 seconds.

Set models.<role>.api_key_env to the environment variable containing the model-provider credential. This adapter does not support MCP, skills, tool policy, or native OpenAI streaming.

import asyncio

from nemo_fabric import EnvironmentConfig, Fabric, FabricConfig, HarnessConfig
from nemo_fabric import InstructionConfig, InstructionsConfig, MetadataConfig
from nemo_fabric import ModelConfig, RuntimeConfig

config = FabricConfig(
    metadata=MetadataConfig(name="mini-swe-agent"),
    harness=HarnessConfig(
        adapter_id="nvidia.fabric.mini-swe-agent",
        resolution="preinstalled",
        settings={"timeout": 30},
    ),
    models={
        "default": ModelConfig(
            provider="nvidia",
            model="nvidia/nemotron-3-nano-omni-30b-a3b-reasoning",
            api_key_env="NVIDIA_API_KEY",
            base_url="https://integrate.api.nvidia.com/v1",
        )
    },
    instructions=InstructionsConfig(
        system=InstructionConfig(content="Work carefully and verify the change.")
    ),
    runtime=RuntimeConfig(max_turns=50, timeout_seconds=1800),
    environment=EnvironmentConfig(provider="local", workspace="/workspace"),
)

async def main():
    return await Fabric().run(
        config, base_dir="/workspace", input="Fix the failing test."
    )


result = asyncio.run(main())

The result includes the submitted final text and API-call usage.

Continue a Task Across Invocations

Use one runtime for ordered invocations that should share mini-SWE-agent conversation history:

async def continue_task():
    async with await Fabric().start_runtime(
        config,
        base_dir="/workspace",
    ) as runtime:
        first = await runtime.invoke(
            input="Inspect calculator.py and identify the bug."
        )
        second = await runtime.invoke(input="Fix it and run the tests.")
    return first, second

For example, the first invocation can produce this conversation:

system: You are a coding agent.
user: Inspect calculator.py and identify the bug.
assistant: I will inspect the file.
assistant action: bash {"command": "cat calculator.py"}
tool: def divide(a, b): return a // b
assistant: The divide function uses // instead of /.
assistant action: bash {"command": "printf 'COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT\\nThe divide function uses integer division.\\n'"}
tool: COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT
      The divide function uses integer division.
exit: Submitted

Before the second model call, the adapter removes the terminal exit message and appends the new input. The retained context resembles the following conversation:

system: You are a coding agent.
user: Inspect calculator.py and identify the bug.
assistant: I will inspect the file.
assistant action: bash {"command": "cat calculator.py"}
tool: def divide(a, b): return a // b
assistant: The divide function uses // instead of /.
assistant action: bash {"command": "printf 'COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT\\nThe divide function uses integer division.\\n'"}
tool: COMPLETE_TASK_AND_SUBMIT_FINAL_OUTPUT
      The divide function uses integer division.
user: Fix it and run the tests.

Assistant messages and output from commands executed through the bash tool are part of the retained model context. Unrelated terminal output is not retained. History lasts only for the same runtime and is discarded when the runtime stops. The adapter does not currently truncate or summarize retained history, so repeated invocations can increase model context usage.

Relay Telemetry

When Relay is enabled, the adapter uses a Relay-specific mini-SWE-agent subclass to emit an Agent scope for each invocation, Function scopes for agent steps, and LLM and tool events for model queries and bash actions. Relay is imported and the subclass is selected only for Relay-enabled runtimes; otherwise the adapter uses the existing retaining agent without Relay instrumentation.

The adapter wraps each invocation with nemo_relay.scope.scope(...). Because mini-SWE-agent does not provide native observability hooks, the adapter-owned subclass explicitly emits nested step, LLM, and bash tool events using Relay handles. This produces a correlated Agent, Function, LLM, and tool scope hierarchy without requiring upstream changes.

Correlation IDs

The top-level mini-swe-agent.request Agent scope stores the NeMo Fabric request, invocation, and runtime IDs, and a caller-supplied session root, as Relay metadata:

NeMo Fabric ID Relay metadata key Meaning
request_id nemo_fabric_request_id Correlates the caller's logical request. A caller can provide the same value when it wants to correlate retries or related processing.
invocation_id nemo_fabric_invocation_id Identifies one concrete invocation attempt. NeMo Fabric assigns a new value to each invocation.
runtime_id nemo_fabric_runtime_id Identifies the longer-lived retained runtime that can process multiple invocations.
RunRequest.relay_session_root nemo_fabric_session_root Groups invocations into one Relay session. A UUID string Relay accepts becomes the propagated Relay root; a UUID request_id stays the parent, otherwise the session root is also the parent. Absent when the value is missing or unusable; the invocation proceeds without an error.

Nested step, LLM, and bash tool events are correlated through the Relay scope hierarchy; they do not repeat these metadata fields.

Install both the harness and Relay:

pip install "nemo-fabric[mini-swe-agent,relay]"

Enable Relay on the configuration:

config.enable_relay()

Relay-enabled runs support configured ATIF, OpenTelemetry, and OpenInference outputs. Runtime.invoke_stream() also provides live ATOF records. Live telemetry uses ordinary adapter invoke() and is independent of native OpenAI streaming, so the adapter's capabilities.streaming value remains false.