Agent Platform Tuning Manager
Inspect, list, and safely cancel Agent Platform GenAI tuning jobs.
The skill clearly separates read-only operations from cancellation, requires explicit typed confirmation before cancellation, and instructs the agent to request missing project or region data and stop after errors. Deductions apply because gcloud login and ADC setup are broad, credential and data-flow handling are not disclosed, dependencies are unpinned, and cancellation has no rollback or recovery mechanism.
The Python SDK examples cover the three stated paths—list, get, and cancel—and specify stopping on certain failures. Static evidence contains no tests, CI, pinned versions, or reproducible results; the 19-digit job-ID assumption may reject valid inputs, and the error guidance conflates permission failures with nonexistent resources. The score therefore remains below the static ceiling.
The skill has a clear scenario and boundary: listing, inspecting, and canceling tuning jobs, while excluding fine-tuning, deployment, and endpoint management. Required inputs are identified. Deductions apply because environmental fit, Chinese-language support, and alternative workflows are not addressed; the core function depends on Google Cloud services whose mainland-China reachability is not discussed.
The document is well organized and includes metadata, safety tiers, setup steps, a decision tree, examples, and failure guidance. The README supplies installation, licensing, contribution, and issue-reporting information. Deductions apply for missing skill versioning, changelog, named maintenance responsibility, dependency pinning, FAQ material, and explicit known limitations.
The examples could directly perform the advertised list, inspect, and cancel operations after placeholders are replaced and authentication is configured. Deductions apply because setup and parameter substitution remain manual, no verified outputs are provided, and static review cannot establish current SDK/API compatibility or comparative benefit over manual use.
The source includes auditable workflows, code samples, safety requirements, repository context, and an Apache-2.0 license. Deductions apply because there are no committed tests, CI coverage, execution logs, or third-party reproduction evidence, so static verifiability is limited.
- Cancellation may lose tuning progress; obtain explicit typed confirmation and verify the project, region, and job ID before execution.
- The google-cloud-aiplatform dependency is unpinned; verify SDK and API compatibility before use.
- gcloud login and ADC modify the local credential environment; the skill does not document credential handling, permission scope, or data flows.
- Network reachability and compliance for Google Cloud services in mainland China are not addressed.
What does this skill do, and when should you use it?
This skill manages GenAI tuning jobs in Agent Platform. It covers listing jobs, retrieving details for a specific job, and canceling a running job only after explicit typed confirmation. It requires a Python virtual environment, Google Cloud authentication with ADC, and the Agent Platform SDK. A valid project ID and region are required; resource or permission errors must stop execution until the user supplies valid details.
It instructs the user to create and activate a Python virtual environment, authenticate with Google Cloud, configure Application Default Credentials, and install google-cloud-aiplatform. It uses GenAiTuningServiceClient.list_tuning_jobs to print each job's name, base model, and state; get_tuning_job to print a specific job's name, base model, state, and tuned model display name; and cancel_tuning_job to request cancellation after confirmation. For 403, 404, INVALID_ARGUMENT, or missing-resource errors, it requires stopping execution and requesting valid project or job details.
- A Google Cloud user wants to see which GenAI tuning jobs are running in a specified project and region.
- A developer has a 19-digit tuning job ID and needs its current state and tuned model display name.
- A user wants to stop a tuning job that is taking too long and accepts that its progress will be lost.
- A developer needs tuning-job management through the Agent Platform Python SDK, separate from model fine-tuning, deployment, or endpoint management.
What are this skill's strengths and limitations?
- Provides concrete SDK workflows for listing, inspecting, and canceling tuning jobs.
- Separates read-only actions from destructive or interruptive cancellation with an explicit confirmation requirement.
- Defines a clear stop-and-request-more-information response for invalid, missing, or inaccessible resources.
- Requires Python 3, the Google Cloud CLI, network access, and the Agent Platform SDK; platform and version testing are not documented.
- The user must supply a project ID and region; the skill does not search arbitrary regions.
- Cancellation loses progress, and no recovery or undo procedure is provided.
- The source does not include a test suite, automation scripts, or retry logic.
How do you install this skill?
Run:
npx skills add google/skills
Select Agent Platform Tuning Management during installation. The README does not document the final local installation directory.
How do you use this skill?
Initialize the environment with:
python3 -m venv ~/tuning_mgr_venv
source ~/tuning_mgr_venv/bin/activate
gcloud auth login
gcloud auth application-default login
pip install google-cloud-aiplatform
Then use a request such as “List tuning jobs for project PROJECT_ID in region REGION” or “Check the status of tuning job JOB_ID.” Keep the virtual environment activated whenever running a Python snippet. For cancellation, first explain that stopping the tuning process will lose progress and ask the user to type “I confirm” or “Yes, cancel it.” Do not provide or execute cancellation code before receiving explicit confirmation in a new turn.
How does this skill compare with similar options?
This skill focuses on inspecting and canceling tuning jobs. Use Agent Platform Tuning for model fine-tuning, Agent Platform Deploy for model deployment, and Agent Platform Endpoint Management for serving-endpoint management.