haipipe-end-develop-local

name	haipipe-end-develop-local
description	Local develop specialist for haipipe-end: runs Stage 5 training on the local machine and produces an Endpoint_Set under 6-EndpointStore/, for dev iteration, smoke tests, and DIY builds. Mostly delegates to /haipipe-nn modelset for training; exists for symmetry with -deploy-local and a uniform develop axis. Reads /haipipe-nn output; writes Endpoint_Sets that endpointset + deploy specialists consume.
argument-hint	[verb] [config_or_modelset] [args...]
allowed-tools	Bash, Read, Write, Edit, Grep, Glob, Skill
metadata	{"version":"0.1.1","last_updated":"2026-07-08","summary":"Local develop specialist for haipipe-end."}

Skill: haipipe-end-develop-local

Local development specialist. Builds an Endpoint_Set on the local machine without any managed training pipeline. Most heavy lifting is done by /haipipe-nn modelset — this skill is a thin orchestration layer that:

1. Invokes the project's local ModelSet pipeline.
2. Packages the resulting trained model into an Endpoint_Set under _WorkSpace/6-EndpointStore/<endpoint_set>/ via the standard haipipe-end-endpointset package flow.

It exists for symmetry with -deploy-local and to give /haipipe-end a uniform develop axis (sagemaker / databricks / local). When local develop is sufficient, prefer this skill over -develop-sagemaker for fast iteration.

Verb axis: dashboard | develop | test | monitor | teardown | review

---

Commands

/haipipe-end-develop-local                                 -> dashboard: locally-built endpoint sets
/haipipe-end-develop-local dashboard                       -> same
/haipipe-end-develop-local develop <modelset_or_config>    -> full local build → Endpoint_Set
/haipipe-end-develop-local test <modelset_or_config>       -> dry-run / smoke (no Endpoint_Set written)
/haipipe-end-develop-local monitor <run_id>                -> tail local logs
/haipipe-end-develop-local teardown <run_id>               -> stop background run, optional cleanup
/haipipe-end-develop-local review <endpoint_set>           -> audit a locally-built Endpoint_Set

<run_id> for local runs is a generated short id; the dashboard lists the local develop registry.

---

Dispatch Table

Verb        Ref                                   Backing call
----------- ------------------------------------- ----------------------------------------
dashboard   ref/concepts.md                       (none — local develop registry)
develop     ref/concepts.md                       Skill("haipipe-nn", "modelset run <args>")
            ../haipipe-end/ref/0-overview.md      then Skill("haipipe-end-endpointset",
                                                              "package <args>")
test        ref/concepts.md                       Skill("haipipe-nn", "modelset test <args>")
monitor     ref/concepts.md                       tail local log file
teardown    ref/concepts.md                       kill pid + optional rm of intermediate dirs
review      ref/concepts.md                       Skill("haipipe-end-endpointset",
                                                              "review <endpoint_set>")

This skill delegates rather than reimplementing. The "local develop" flow is just nn modelset run followed by endpointset package — both are already first-class skills, so this wrapper just sequences them.

---

Step-by-Step Protocol

Step 0: Read ref/concepts.md for local develop conventions (pid file layout, log paths, registry location).

Step 1: Parse args. Required arg per verb: develop / test: <modelset_or_config> monitor / teardown: <run_id> review: <endpoint_set_name>

Step 2: For develop: a) Dispatch: Skill("haipipe-nn", args="modelset run ") Capture status / artifacts. If status != ok, abort and return. b) Dispatch: Skill("haipipe-end-endpointset", args="package <endpoint_set_args>") Capture status / artifacts. c) Record (run_id, modelset, endpoint_set_path) in the local registry.

    For `test`:
      Dispatch:    Skill("haipipe-nn", args="modelset test <modelset>")
      (No Endpoint_Set written — fast iteration.)

    For `review`:
      Dispatch:    Skill("haipipe-end-endpointset", args="review <endpoint_set>")

Step 3: Emit the structured tail (composing the chained tails):

status:    ok | blocked | failed   (failed if either delegate failed)
summary:   2-3 sentences; mention which step succeeded / failed
artifacts: [run_id, modelset run output, endpoint_set path]
next:      suggested next command (typically /haipipe-end deploy local <endpoint_set>)

---

Why this skill is thin

The actual training work lives in /haipipe-nn and the actual packaging lives in /haipipe-end-endpointset. This skill exists for two reasons:

1. Symmetry. -deploy-local exists alongside -deploy-sagemaker / -databricks / -mlflow. For the -develop- axis to be uniform, -develop-local must also exist.
2. Uniform routing. /haipipe-end develop sagemaker <args> and /haipipe-end develop local <args> should both work without the orchestrator special-casing local. Having this specialist gives the develop dispatch table a uniform shape.

If you find yourself adding non-trivial local-only logic here, that's a signal it probably belongs in /haipipe-nn or /haipipe-end-endpointset instead. Keep this skill thin.

---

Target Scope

Owns:

• The nn modelset → endpointset package sequencing for local builds
• Local develop registry (which builds were produced when, where they live)
• Local pid / log management for background develop runs

Does NOT own:

• ANY actual training logic — that's /haipipe-nn modelset.
• ANY actual Endpoint_Set packaging logic — that's /haipipe-end-endpointset.
• Cloud / managed training — see -develop-sagemaker (or future -develop-databricks).

If a develop fails because of a training issue, escalate to /haipipe-nn modelset review. If it fails because of a packaging issue, escalate to /haipipe-end-endpointset review. This skill should not patch either.