training-sprint-loop

name	training-sprint-loop
description	Run a repeatable training sprint loop: implement, validate, commit, push, and CI check with clear evidence and scoped commits.
argument-hint	Task slice, validation scope, and commit intent
user-invocable	true
disable-model-invocation	false

Training Sprint Loop

Inherits from task-core-loop.

Use this skill for short, repeatable training-focused delivery rounds where each round ends in a pushed checkpoint and CI status update.

Use When

• You want commit/push/review/fix/repeat cadence.
• You are iterating on training harness behavior, telemetry contracts, or workbench integration.

Inputs

• Change slice and intended outcome.
• Validation scope (focused tests, strict GPU run, or both).
• Commit scope (files that must be included/excluded).

Procedure Additions

1. Apply all task-core-loop steps first.
2. Add strict or focused training validation for the changed slice.
3. Push and capture CI status with conservative checks.
4. Record run-artifact evidence (status.json, benchmark.json, event traces) when relevant.

Conservative CI Mode

• Prefer one gh run list status refresh per round.
• Avoid frequent polling loops unless user explicitly requests live watch.
• If runs are queued/in-progress, continue next safe local slice instead of aggressive CI polling.

Done Criteria

• Changes applied and validated for the slice.
• Commit pushed with clean scope.
• CI state captured (queued/in-progress/success/failure).
• Tracker/memory notes updated with evidence.