Opencode plugin artifacts error

[0byte-coding]

I wanted to try out the "plugin": ["@mlflow/opencode"] since todays mlflow release 🎉 but it didnt work out that great. The trace still appears in mlflow but opencode throws an error.

Not sure what the issue is, I used the default artifact location which was basic_auth.db and had the errors and hoped when switching to rustfs s3 setup that would fix it but im still having the same issue. I used the --default-artifact-root option before which apparently proxies directly to the artifact store in case thats relevant. Hoping someone can help me out with this.

How I ran opencode:

export MLFLOW_TRACKING_URI=https://mymlflowdomain.com
export MLFLOW_EXPERIMENT_ID=69
export MLFLOW_TRACKING_USERNAME=admin
export MLFLOW_TRACKING_PASSWORD=password1234

opencode

The command im currently using for mlflow:v3.11.1 docker image:

/bin/sh
      -c
      pip install --no-cache-dir psycopg2-binary boto3 mlflow[auth]
      mlflow server \
        --backend-store-uri "${MLFLOW_BACKEND_STORE_URI}" \
        --artifacts-destination "${MLFLOW_ARTIFACTS_DESTINATION}" \
        --serve-artifacts \
        --host "${MLFLOW_HOST}" \
        --port "${MLFLOW_PORT}" \
        --allowed-hosts "*" \
        --cors-allowed-origins "*" \
        --app-name basic-auth \
        --workers 1

The stderr happens when it finishes its last message from what I can tell.
What it writes to stderr (sorry if it looks ugly, not sure how to get normal looking stderr output):

�[0m�[31mFailed to export trace tr-98db0ebbf9f35676033b81aed497e047: �[0m�[1m84 |�[0m      * @param fileName The file name to append (e.g., �[0m�[32m"traces.json"�[0m)
�[0m�[1m85 |�[0m      * @returns The resolved HTTP endpoint URL
�[0m�[1m86 |�[0m      */
�[0m�[1m87 |�[0m     resolveArtifactUri(artifactUri, fileName) {
�[0m�[1m88 |�[0m         �[0m�[35mconst�[0m baseApiPath = �[0m�[32m'/api/2.0/mlflow-artifacts/artifacts'�[0m�[0m�[2m;�[0m
�[0m�[1m89 |�[0m         �[0m�[35mconst�[0m url = �[0m�[35mnew�[0m �[0m�[1mURL�[0m(artifactUri)�[0m�[2m;�[0m
                         �[31m�[1m^�[0m
�[0m�[31mTypeError�[0m�[2m:�[0m �[1m"/mlflow/artifacts/38/traces/tr-98db0ebbf9f35676033b81aed497e047/artifacts" cannot be parsed as a URL.�[0m
 code�[0m�[2m:�[0m �[32m"ERR_INVALID_URL"�[0m

�[0m      �[2mat �[0m�[0m�[1m�[3mresolveArtifactUri�[0m�[2m (�[0m�[0m�[36m/home/dev/.cache/opencode/packages/@mlflow/opencode@latest/node_modules/@mlflow/core/dist/clients/artifacts/mlflow.js�[0m�[2m:�[0m�[33m89�[0m�[2m:�[33m21�[0m�[2m)�[0m
�[0m      �[2mat �[0m�[0m�[1m�[3muploadTraceData�[0m�[2m (�[0m�[0m�[36m/home/dev/.cache/opencode/packages/@mlflow/opencode@latest/node_modules/@mlflow/core/dist/clients/artifacts/mlflow.js�[0m�[2m:�[0m�[33m37�[0m�[2m:�[33m34�[0m�[2m)�[0m
�[0m      �[2mat �[0m�[0m�[1m�[3muploadTraceData�[0m�[2m (�[0m�[0m�[36m/home/dev/.cache/opencode/packages/@mlflow/opencode@latest/node_modules/@mlflow/core/dist/clients/client.js�[0m�[2m:�[0m�[33m86�[0m�[2m:�[33m36�[0m�[2m)�[0m
�[0m      �[2mat �[0m�[0m�[1m�[3mexportTraceToBackend�[0m�[2m (�[0m�[0m�[36m/home/dev/.cache/opencode/packages/@mlflow/opencode@latest/node_modules/@mlflow/core/dist/exporters/mlflow.js�[0m�[2m:�[0m�[33m167�[0m�[2m:�[33m32�[0m�[2m)�[0m
�[0m
�[0m�[31mFailed to export trace tr-98db0ebbf9f35676033b81aed497e047: �[0m�[1m84 |�[0m      * @param fileName The file name to append (e.g., �[0m�[32m"traces.json"�[0m)
�[0m�[1m85 |�[0m      * @returns The resolved HTTP endpoint URL
�[0m�[1m86 |�[0m      */
�[0m�[1m87 |�[0m     resolveArtifactUri(artifactUri, fileName) {
�[0m�[1m88 |�[0m         �[0m�[35mconst�[0m baseApiPath = �[0m�[32m'/api/2.0/mlflow-artifacts/artifacts'�[0m�[0m�[2m;�[0m
�[0m�[1m89 |�[0m         �[0m�[35mconst�[0m url = �[0m�[35mnew�[0m �[0m�[1mURL�[0m(artifactUri)�[0m�[2m;�[0m
                         �[31m�[1m^�[0m
�[0m�[31mTypeError�[0m�[2m:�[0m �[1m"/mlflow/artifacts/38/traces/tr-98db0ebbf9f35676033b81aed497e047/artifacts" cannot be parsed as a URL.�[0m
 code�[0m�[2m:�[0m �[32m"ERR_INVALID_URL"�[0m

�[0m      �[2mat �[0m�[0m�[1m�[3mresolveArtifactUri�[0m�[2m (�[0m�[0m�[36m/home/dev/.cache/opencode/packages/@mlflow/opencode@latest/node_modules/@mlflow/core/dist/clients/artifacts/mlflow.js�[0m�[2m:�[0m�[33m89�[0m�[2m:�[33m21�[0m�[2m)�[0m
�[0m      �[2mat �[0m�[0m�[1m�[3muploadTraceData�[0m�[2m (�[0m�[0m�[36m/home/dev/.cache/opencode/packages/@mlflow/opencode@latest/node_modules/@mlflow/core/dist/clients/artifacts/mlflow.js�[0m�[2m:�[0m�[33m37�[0m�[2m:�[33m34�[0m�[2m)�[0m
�[0m      �[2mat �[0m�[0m�[1m�[3muploadTraceData�[0m�[2m (�[0m�[0m�[36m/home/dev/.cache/opencode/packages/@mlflow/opencode@latest/node_modules/@mlflow/core/dist/clients/client.js�[0m�[2m:�[0m�[33m86�[0m�[2m:�[33m36�[0m�[2m)�[0m
�[0m      �[2mat �[0m�[0m�[1m�[3mexportTraceToBackend�[0m�[2m (�[0m�[0m�[36m/home/dev/.cache/opencode/packages/@mlflow/opencode@latest/node_modules/@mlflow/core/dist/exporters/mlflow.js�[0m�[2m:�[0m�[33m167�[0m�[2m:�[33m32�[0m�[2m)�[0m
�[0m

Also im a bit surprised that mlflow does not track all too much? For example the model thinking is completely excluded which seems to defeat the point imo? Would also expect it to log the initial systemprompt/tools.

I would not really mind resetting my database if that would fix it but I would like to export and then import my experiments back, I heard that they can be archived to s3 or so? could someone point me in the direction on how to do so? Thanks in advance

[0byte-coding]

I cleaned up my deployment and set it up again, this time storing authentication details in the postgres db and using s3 for artifacts and the opencode plugin doesnt error anymore.

I see that tool calls are being captured in the traces but the toolcalls dont show up on the overview page?

[SahilKumar75]

The tool calls are captured as child spans of type TOOL inside the trace, so they show up in the trace detail view in the span tree. The overview page shows a high-level summary of the trace (request/response, token usage, latency) rather than the individual spans. To see the tool calls, open the trace and look at the span tree on the left — each tool invocation will appear as a child span under the LLM span.

[smqd19]

Hey there, sorry to hear you're running into issues with the @mlflow/opencode plugin on the latest MLflow release. I've worked with MLflow extensively in production for tracking and artifact storage, so let's dig into this. From your setup, it looks like you're using a custom artifact destination and backend store, which might be where the conflict arises—especially since you mentioned switching to RustFS S3 and still seeing errors.

First off, can you share the exact error message from stderr? That’ll help narrow down whether this is a permissions issue, a compatibility problem with the plugin, or something else. We’ve seen similar issues in the past when artifact stores (like S3) have mismatched credentials or when the --default-artifact-root flag doesn’t align with the plugin’s expectations. Also, since you're running MLflow in a Docker container with basic-auth, double-check that the environment variables for auth are correctly passed to the opencode process—sometimes Docker env propagation can be tricky.

As a quick test, try running opencode with a local artifact store (like a simple file path) instead of S3 to isolate if the issue is with the remote setup. Here's a minimal command to try that:

export MLFLOW_TRACKING_URI=https://mymlflowdomain.com
export MLFLOW_EXPERIMENT_ID=69
export MLFLOW_TRACKING_USERNAME=admin
export MLFLOW_TRACKING_PASSWORD=password1234
export MLFLOW_DEFAULT_ARTIFACT_ROOT=./artifacts
opencode

Let me know if that changes anything or if you’ve got the error logs handy—I’m happy to take a closer look. Also, are you on the latest version of the opencode plugin? We’ve had to pin specific plugin versions in the past when MLflow updates break compatibility.

[SahilKumar75]

The error is straightforward the artifactUri returned by your MLflow server is a relative path (/mlflow/artifacts/38/traces/...) rather than an absolute URL, and the @mlflow/opencode plugin's resolveArtifactUri method passes it directly into new URL(...), which throws ERR_INVALID_URL on relative paths.

Why it happened with your original setup: When you use --default-artifact-root and serve behind a reverse proxy, MLflow sometimes returns the artifact URI as a relative path rather than a fully-qualified URL (e.g. https://mymlflowdomain.com/mlflow/artifacts/...). The plugin doesn't handle this case.

What fixed it for you (your own follow-up): Switching to a clean setup with S3 artifacts likely caused MLflow to return a proper s3://... or https://... URI, which new URL() can parse without issue.

For others hitting this: A few things to check:

1. What does your artifact URI actually look like? You can verify by calling the MLflow REST API directly:

   curl -u admin:password1234 https://mymlflowdomain.com/api/2.0/mlflow/experiments/get?experiment_id=69
   

   Check the artifact_location field — if it's a relative path or starts with /, that's the problem.

2. If behind a reverse proxy, make sure MLflow can resolve its own public base URL. Setting MLFLOW_TRACKING_URI on the server side or passing --host + a proper public URL helps MLflow construct absolute artifact URIs.

3. This looks like a bug in @mlflow/opencode — resolveArtifactUri should handle relative paths by prepending the tracking URI base. Worth filing a separate issue against the plugin.

Regarding your follow up about tool calls not showing on the overview page that's likely expected behavior; the overview typically shows span summaries, not individual tool call details. The full tool call data lives in the trace detail view. If you're not seeing it there either, could be a UI rendering issue with that version.