Each pipeline command should suggest the next step with the exact command to run

Summary

Every command in the ModelKit pipeline completes and exits without telling the developer what to do next. winml inspect is a dead end — the developer sees model info but has to guess the next command, its flags, and the right values. Every stage should close with a "Next step" hint that carries forward the model ID, device, and EP context already established.

Motivation

ModelKit has two distinct pipelines with a clear step order, but nothing in the CLI output surfaces either of them:

Pipeline A — HuggingFace model, high-level:

inspect → config → build → perf → eval

Pipeline B — Low-level / advanced:

# From a HuggingFace model:
inspect → export → analyze → optimize → quantize → compile → perf → eval

# From an existing ONNX file (pipeline starts at analyze):
analyze → optimize → quantize → compile → perf → eval

A developer who runs winml inspect dslim/bert-base-NER sees model info but receives no signal about which pipeline applies or what to run next.

Proposed "Next:" hints per command

Pipeline A — HuggingFace, high-level

winml inspect <hf_model> →

  Next:  winml config dslim/bert-base-NER --ep qnn --device npu
         (for step-by-step control: winml export dslim/bert-base-NER -o ./build/)

winml config →

  Next:  winml build dslim/bert-base-NER -c dslim-bert-base-NER_config.json -o ./build/

winml build →

  Next:  winml perf ./build/model.onnx --device npu

---

Pipeline B — Low-level, HuggingFace model

winml export <hf_model> →

  Next:  winml analyze ./build/model.onnx --ep qnn

winml analyze →

  Next:  winml optimize ./build/model.onnx -o ./build/

winml optimize →

  Next:  winml quantize ./build/optimized.onnx -o ./build/

winml quantize →

  Next:  winml compile ./build/quantized.onnx --ep qnn -o ./build/

winml compile →

  Next:  winml perf ./build/compiled.onnx --device npu

---

Pipeline B — Low-level, ONNX file input (pipeline starts at analyze)

winml analyze model.onnx →

  Next:  winml optimize model.onnx -o ./build/

(same hints from optimize onward as above)

---

Shared tail

winml perf →

  Next:  winml eval ./build/model.onnx --device npu

Design Notes

• ONNX input skips inspect and export. For ONNX file inputs Pipeline B starts at analyze — no inspect or export step is shown.
• inspect only applies to HF models. It surfaces Pipeline A (config) as the primary path and Pipeline B (export) as the advanced alternative.
• Carry context forward. Hints use the actual model ID, output path, EP, and device from the current invocation — never generic placeholders.
• Respect skipped stages. If a stage was skipped (e.g. quantize skipped because model is already QDQ), the hint points to the next applicable step.
• --no-hints flag. Suppress next-step hints for CI / scripted use.
• winml flow relationship. The winml flow guided pipeline (#442) is the automated version of these same sequences — the hints are the manual-mode equivalent and share the same step-ordering logic.

Acceptance Criteria

Pipeline A (HF, high-level):

• winml inspect <hf_model> hints to winml config with Pipeline B option noted
• winml config hints to winml build with the generated config file path
• winml build hints to winml perf with the final artifact path

Pipeline B (HF model, low-level):

• winml export hints to winml analyze with the exported ONNX path
• winml analyze hints to winml optimize
• winml optimize hints to winml quantize with the optimized ONNX path
• winml quantize hints to winml compile with the quantized ONNX path
• winml compile hints to winml perf with the compiled artifact path

Pipeline B (ONNX file input):

• winml analyze <onnx_file> hints to winml optimize (no inspect/export shown)
• Remaining steps (optimize → quantize → compile → perf) hint identically to HF path

Shared:

• winml perf hints to winml eval
• All hints use actual runtime values — model ID, output path, EP, device
• --no-hints flag suppresses the hint block on any command
• Skipped stages are omitted from hints (next applicable step shown)

Related

• #442 — winml flow guided pipeline (automated version of both sequences)
• #447 — positional model argument (hints should use positional form for readability)