sir-convert-a-lot-colab-hemma

name: sir-convert-a-lot-colab-hemma description: >- Workflow skill for Sir Convert-a-Lot Colab notebook orchestration and Hemma-backed preprocessing iteration. Use when the task involves portable Colab Qwen slices, notebook-backed row-processing, syncing code between Colab and Hemma, or deciding where code edits and git operations should happen for a Hemma-executed lane. Trigger especially when a Colab notebook is only the operator shell and the real execution, artifacts, or canonical repo state live on Hemma.

Sir Convert-a-Lot Colab Hemma

Purpose

Keep Colab thin, keep Hemma authoritative, and keep repo iteration close to the runtime that will actually execute the work.

Use this skill together with:

• .codex/skills/sir-convert-a-lot-qwen-finetuning/SKILL.md
• .codex/skills/sir-convert-a-lot-devops-hemma/SKILL.md
• docs/runbooks/runbook-qwen3-swedish-finetuning-on-hemma-and-colab.md

Core Rule

If the execution lane is Hemma-backed, do the repo edits on Hemma and push from Hemma.

Apply that rule to:

• Python/module changes for Hemma-run preprocessing or training
• notebook edits when the notebook is loaded from the Hemma repo
• dependency changes needed by the Hemma execution lane
• commits and pushes for Hemma-owned work

Do not default to editing locally first and then trying to replay the same changes on Hemma. That adds latency, risks drift, and makes notebook/runtime debugging slower.

Workflow

1. Classify the lane.
   • Hemma-backed lane: Hemma is the execution truth for preprocessing, training, artifact generation, or the notebook source tree.
   • Colab-only lane: Colab is disposable compute, but the notebook still orchestrates committed repo commands.
2. Decide where repo edits belong.
   • If Hemma-backed lane, edit in the Hemma repo clone and push from Hemma.
   • If purely local planning/docs work with no Hemma execution dependency, local edits are acceptable.
3. Keep notebooks thin.
   • The notebook may install deps, define paths, and invoke committed CLI or module surfaces.
   • Do not move archive resolution, row selection, or preprocessing logic into notebook-only cells.
4. Sync in the safe direction.
   • For Hemma-backed work: edit on Hemma, commit on Hemma, push from Hemma, then pull locally if a local clone needs the result.
   • For local-only docs or exploratory notes: commit locally only if Hemma is not the execution owner for the same slice.

Colab Rules

• Treat Colab as an operator shell around committed repo surfaces.
• Start from a Hemma-issued portable slice or other committed artifact set.
• Stage only required files.
• Localize the slice through repo-owned commands when needed.
• Run canonical Task 103 row-processing against committed manifests.
• Preserve restartability through repo-owned resume semantics, not notebook patches.
• If Colab persists run state into Google Drive and the Drive connector is authenticated, inspect status.json, spool JSON, and row-processing logs directly through Drive before asking the user for manual notebook commands.
• Treat Drive-backed Colab artifacts as part of the operational surface for this lane, not as inaccessible notebook-only state.
• If the user provides a direct Drive file or folder link, use that id/path first. Do not start with broad search when the precise artifact is already known.
• Expect connector quirks:
  • canonical folder URLs work better than /u/1/ variants
  • metadata lookup by id may succeed even when content fetch or search is weak
  • top-level run-root folders may expose logs immediately while deeper spool or status artifacts still need direct links

Notebook Hygiene

• Expect notebook metadata drift such as kernel display-name changes.
• Do not treat metadata-only notebook edits as meaningful implementation.
• If a Hemma notebook clone picks up metadata-only changes during execution, discard them before pull/merge unless the metadata was intentionally changed.

Command Preference

• Use canonical wrappers for Hemma work:
  • pdm run run-hemma -- <command> [args]
• Prefer committed scripts or module surfaces over ad hoc shell payloads.
• Use merge-only git workflow. Never rebase.

Escalation Guide

Pause and confirm with the user when:

• Hemma has unrelated uncommitted repo changes that are not metadata-only
• the Hemma worktree has diverged from local in ways that would affect runtime
• the fix would require discarding anything beyond obvious notebook metadata
• a change would move logic from repo code back into notebook cells

Success State

• Hemma remains the source of truth for Hemma-backed execution lanes.
• Colab remains simple enough to "press run."
• Drive-backed Colab artifacts are inspected directly when connector access is available.
• The repo history reflects the environment that actually executed the change.