• Home
  • Features
  • Pricing
  • Docs
  • Announcements
  • Sign In

kubeflow / mcp-server / 44363181166
95%

Build:
DEFAULT BRANCH: main
Ran 18 Nov 2026 05:52PM UTC
Jobs 3
Files 57
Run time 1min
Badge
Embed ▾
README BADGES
x

If you need to use a raster PNG badge, change the '.svg' to '.png' in the link

Markdown

Textile

RDoc

HTML

Rst

18 Nov 2026 05:50PM UTC coverage: 94.604% (+0.5%) from 94.074%
44363181166

push

github

web-flow
feat(pipelines): create pipelines  and manage experiments New tools  (#561)

* feat(pipelines): add create_pipeline, delete_pipeline and experiment tools

The pipelines client could only run pipelines that were already registered,
so an agent could not turn a request like "train this model on this dataset"
into a pipeline, and experiments had no tools at all.

create_pipeline takes KFP v2 code written by the agent. The preview checks
the code's AST without running it: KFP runs the module top level and the
@dsl.pipeline body while compiling, so only kfp/typing imports, literal
constants, @dsl.component functions and one @dsl.pipeline that wires steps
together are accepted. Component bodies run in the cluster and stay free.
On confirm the code is compiled in a separate python -I process with a
stripped environment and a timeout, the YAML is checked (images, GPUs,
size) and uploaded unchanged with PipelinesClient.upload_pipeline, so the
usual KFP header and platform spec are kept.

Also adds delete_pipeline and list/get/create/delete_experiment. The delete
tools are platform-admin only because KFP pipelines and experiments carry
no MCP ownership label.

Adds the pipelines extra (also in all), so kfp installs with the server.
The lock now resolves kubernetes 30.1.0 because kfp 2.17 needs <31. The
Hugging Face dataset checks move to common/huggingface.py so trainer and
pipelines share them; trainer behaviour is unchanged.

Tested end to end on EKS with KFP 2.17.3.

Fixes #560

Signed-off-by: pratik naik <naikp0241@gmail.com>

* fix(pipelines): read KFP v2 launcher logs by task name

On KFP 2.17, get_pipeline_logs(task="evaluate-model") answered "pod no
longer exists" even though the step's pod was running. KFP records each
launcher pod under a task called "executor", so the user's task name only
pointed at a DAG node that has no pod. Without a task filter the 10k log
budget was also split across those pod-less nodes, so each real pod got
about 1k character... (continued)

5207 of 5504 relevant lines covered (94.6%)

2.84 hits per line

Coverage Regressions

Lines Coverage ∆ File
47
90.98
-0.44% trainer/api/training.py
7
96.24
0.4% pipelines/api/monitoring.py
5
95.28
2.98% pipelines/types/__init__.py
Jobs
ID Job ID Ran Files Coverage
1 python-3.11 - 44363181166.1 18 Nov 2026 05:52PM UTC 57
94.6
GitHub Action Run
2 python-3.12 - 44363181166.2 18 Nov 2026 05:52PM UTC 57
94.6
GitHub Action Run
3 python-3.10 - 44363181166.3 18 Nov 2026 05:52PM UTC 57
94.6
GitHub Action Run
Source Files on build 44363181166
  • Tree
  • List 57
  • Changed 10
  • Source Changed 0
  • Coverage Changed 10
Coverage ∆ File Lines Relevant Covered Missed Hits/Line
  • Back to Repo
  • Github Actions Build #44363181166
  • 2c3e6896 on github
  • Prev Build on main (#44347470662)
  • Next Build on main (#44467871744)
  • Delete
STATUS · Troubleshooting · Open an Issue · Sales · Support · CAREERS · ENTERPRISE · START FREE TRIAL · SCHEDULE DEMO
ANNOUNCEMENTS · TWITTER · TOS & SLA · Supported CI Services · What's a CI service? · Automated Testing

© 2026 Coveralls, Inc