haipipe-task-for-endpoint

name	haipipe-task-for-endpoint
description	endpoint task-folder specialist: scaffolds AND executes {NN}_<name>/ task-folders that package a trained ModelInstance_Set into a deployable Endpoint_Set (Stage 6) via c_endpoint_nb.py. Called by /haipipe-task when task-type=endpoint. Cross-references /haipipe-end for Fn authoring and deploy targets.
argument-hint	[project_id] [group] [task-name]
allowed-tools	Bash, Read, Write, Edit, Grep, Glob, Skill
metadata	{"version":"0.2.3","last_updated":"2026-07-08","summary":"endpoint task-folder specialist (scaffold + execute + deploy prep)."}

Skill: haipipe-task-for-endpoint

Scaffolds an endpoint task-folder (default C-series; letters are project-specific). Takes a trained ModelInstance_Set (Stage 5, with examples) and packages it into a self-contained Endpoint_Set (Stage 6) using Endpoint_Pipeline.

The task .py is an exact copy of code/scripts/haistepnb/c_endpoint_nb.py. CONFIG is overridden at runtime by papermill, NOT by editing the file.

Invocation modes: interactive (human steers) OR headless (haipipe-task-creator-agent calls this skill during Phase 2: Build).

Position in the ABC pipeline

A-series  →  B-series  →  C-series  →  Deploy
data         model         endpoint     (local / databricks / sagemaker)
(aa/ab/ac/ad)  (b_model_nb)  (c_endpoint_nb)

haipipe-task-for-data   haipipe-task-for-fit   haipipe-task-for-endpoint
                                                       ↓
                                               /haipipe-end deploy <target>

What this scaffolds

tasks/C{NN}_{group_name}/
  └── {NN}_{task_name}/
      ├── 1_{task_name}.py           ← exact copy of c_endpoint_nb.py
      ├── configs/
      │   └── run_{task_name}.yaml   ← endpoint config (5 Fn names + model ref)
      ├── runs/
      │   └── run_{task_name}.sh     ← papermill runner
      ├── results/                   ← (created at runtime)
      └── notebooks/                 ← (created at runtime)

Config YAML structure

Seeded from ref/config-seed.yaml:

# Source model (Stage 5)
modelinstance_name: "<name>"
modelinstance_version: "<version>"

# Target endpoint (Stage 6)
endpoint_name: "<endpoint_name>"
endpoint_version: "<version>"

# 5 Inference function names (must exist in code/haifn/fn_endpoint/)
MetaFn:      "<MetaFnName>"
TrigFn:      "<TrigFnName>"
PostFn:      "<PostFnName>"
Src2InputFn: "<Src2InputFnName>"
Input2SrcFn: "<Input2SrcFnName>"

deployment_config:
  environment: "dev"
  platform: "local"     # local | databricks | sagemaker

Pipeline flow (c_endpoint_nb.py steps)

[1] Load ModelInstance_Set (with examples from ExampleConfig)
[2] Endpoint_Pipeline.run()
    → loads examples from ModelInstance_Set.examples_data
    → Src2InputFn(ProcName_to_ProcDf) → payload.json per example
    → packages Endpoint_Set
[3] Save to 6-EndpointStore/
[4] Verify examples have payload.json
[5] Test inference with sample payloads
[6] Package for deployment (.tar.gz)

Output structure:

6-EndpointStore/{endpoint_name}/   (canonical layout: ../../haipipe-end/ref/0-overview.md)
├── model/                    ← copied from ModelInstance_Set
├── code/                     ← codebase snapshot (haipipe, hainn, haifn)
├── examples/                 ← from ModelInstance_Set + payload.json
│   ├── example_000/
│   │   ├── ProcName_to_ProcDf/   real source data (parquet)
│   │   ├── payload.json          generated by Src2InputFn
│   │   ├── example_info.json     case identifiers
│   │   ├── prediction_results.json
│   │   └── df_case_example.json
│   └── ...
├── meta.json                 ← MetaFn output (name mappings + metadata)
└── manifest.json             ← config + full lineage chain

Prerequisites

1. Stage 5 complete with examples:

   • ModelInstance_Set must have examples_data (from ExampleConfig in training config)
   • Each example has ProcName_to_ProcDf/ (real patient source data as parquet)

2. 5 Fns authored in code/haifn/fn_endpoint/:

   • Build via the endpoint fn_develop task folder’s {a1,b1,c1,d1,e1}_build_*.py (legacy: code-dev/1-PIPELINE/6-Endpoint-WorkSpace/)
   • OR use /haipipe-end <fn-type> design interactively

3. ExampleConfig in B-series training config:

   ExampleConfig:
     enabled: true
     ExampleFnName: <name>
     num_examples: 3
   

Execution

CLI (A layer):

python code/scripts/haistepcli/endpoint.py \
    --config examples/.../C01_.../configs/run_endpoint_*.yaml

Notebook (B layer — papermill):

papermill 1_endpoint_*.ipynb output.ipynb \
    -p CONFIG examples/.../C01_.../configs/run_endpoint_*.yaml

After packaging → deploy

/haipipe-end deploy local <endpoint>       ← FastAPI on localhost
/haipipe-end deploy databricks <endpoint>  ← Databricks Model Serving
/haipipe-end deploy sagemaker <endpoint>   ← AWS SageMaker

Latency profiling (formerly this skill's scope) is now a verb:

/haipipe-end profile <endpoint>

Cross-references

• /haipipe-end — Fn authoring, endpointset verbs, deploy targets
• /haipipe-task-for-fit — B-series training (produces the ModelInstance_Set)
• /haipipe-task-for-data — A-series data pipeline (produces the AIDataSet)
• code/scripts/haistepnb/c_endpoint_nb.py — the template
• code/scripts/haistepcli/endpoint.py — the CLI entry point
• ref/notebook-templates.md (in haipipe-task-for-data) — full ABC template guide

Scaffold flow

1. Identify project + task-group (letter C).
2. Collect metadata (NN, name, model ref, 5 Fn names, _meta).
3. Copy c_endpoint_nb.py to task folder (exact copy).
4. Seed config from ref/config-seed.yaml.
5. Create runs/run_{task_name}.sh from template.
6. Emit return contract.

MUST NOT

• Edit the template .py — it is an exact copy of c_endpoint_nb.py.
• Put project configs in code/scripts/haistepconfig/ — those are reference templates only.
• Skip ExampleConfig in training — without examples, no payload.json.
• Create README.md.

Return contract

status:     ok | blocked | failed
task_folder: <path to scaffolded task folder>
run_name:   run_{task_name}
files:      [1_{task_name}.py, configs/run_{task_name}.yaml, runs/run_{task_name}.sh]
next:       run the task, then /haipipe-end deploy <target>