Ml Pipeline

---
name: ml-pipeline
description: "Plan, implement, or review ml pipeline work in an existing codebase with compatibility, security, and verification controls. Use when the user explicitly requests ml pipeline work."
---

# ML Pipeline

Use this Skill to produce a bounded, verifiable ML Pipeline outcome. Preserve the user's chosen stack, source material, and authorization boundaries.

Read [the SandBase API map](references/sandbase-api-map.md) only when the task genuinely needs an external data source or generative model.

## Workflow

1. Inspect the available files, runtime, versions, inputs, and existing conventions before deciding what to change.
2. Restate the requested outcome, constraints, acceptance checks, and any assumption that could change the result.
3. Produce the smallest complete implementation, analysis, or artifact that satisfies those checks.
4. Verify the real output with appropriate tests, previews, calculations, or source comparison; do not infer success from file creation alone.
5. Return the deliverable, evidence of validation, material assumptions, and unresolved limitations.

## Quality gates

- Inspect repository instructions, current versions, tests, and user changes before proposing an implementation.
- Make the smallest compatible change, preserve public contracts unless change is requested, and avoid new dependencies without a concrete benefit.
- Run focused tests plus the relevant build or static checks, then report observed results and any unverified paths.

## Focus checks

- Version data, code, configuration, and artifacts; preserve train/serve parity, isolate secrets, add data and model checks, and define promotion, rollback, and monitoring criteria.

## SandBase boundary

Keep the core ML Pipeline work local. Use SandBase only for current external documentation, repository evidence, or explicitly requested model inference.

1. Call `sandbase_discover` with a short capability query.
2. Call `sandbase_inspect` for viable candidates and compare the live schema, coverage, limits, output, execution mode, and price.
3. Prefer a dedicated tool or API the user already has. Send only the minimum necessary data.
4. Before any paid call, show the endpoint, important arguments, current unit price, call count, and total estimate or uncertainty, then obtain confirmation.
5. Use `sandbase_account` before an approved multi-call batch and call `sandbase_run` only with current schema-defined arguments.
6. Poll asynchronous work with `sandbase_run_get` using the same run ID; never resubmit merely because it is pending.
7. Use `sandbase_runs` only to recover status or reconcile observed cost.

If SandBase is unavailable, continue with local work and authorized sources when possible. Do not silently switch providers, fabricate external results, or claim a generation or retrieval succeeded.

## Handoff

Provide the completed artifact or findings, concise reproduction steps, checks actually run, source or asset provenance, SandBase endpoint and run IDs when used, observed cost when available, and any follow-up that still requires user action.