Extract git-repo ingestion into a standalone /ingest-course-repo skill

[Carr1005]

Reverses the coupling introduced in #2 by extracting it rather than reverting it. /generate-spec goes back to being purely materials → spec.md; repo ingestion becomes its own skill.

Not a git revert of d62df53 — #3/#6/#7 rewrote generate-spec/SKILL.md afterwards, and a revert would clobber the guide-deference wording. The script moves via git mv, so its history follows.

Why

/generate-spec had to carve out an exception in its own ## Don't list — "Don't write to materials/… except the Step 2 ingest script" — an invariant with an exception inside the skill meant to honour it. It also made a single-purpose skill do two jobs, which nothing else in the loop does (/new-course deliberately doesn't generate the spec).

After: /generate-spec has a flat Don't write to builds/, evals/, or materials/, and /ingest-course-repo is the sole writer into materials/.

Two bugs found and fixed

Silent clobber. The --out path #2 prescribed (materials/notebooks/<course>-context.md) is byte-identical to the name of a hand-downloaded context dump — confirmed against the eval fixtures — and out.write_text overwrites without asking. Output now goes to materials/notebooks/from-repo/, so both coexist.

The target repo couldn't be ingested at all. deeplearningai-eng/courses is a private 3.4 GB monorepo, 262 root entries, one course_<id>/ per course, git-LFS-backed. find_notebooks/find_helpers rglob from the clone root with no scoping, so pointing the skill at it would have cloned 3.4 GB and merged ~260 courses into one context file.

ingest_repo.py gains an optional --subdir that scopes discovery and switches the clone to --filter=blob:none --sparse with GIT_LFS_SKIP_SMUDGE=1:

	before	after
course_10 ingest	3.4 GB clone, ~260 courses merged	8.7 MB, ~8s, 5 notebooks

Omitted, behaviour is byte-identical to today — standalone single-course repos are unaffected.

Interface

/ingest-course-repo https://github.com/deeplearningai-eng/courses/tree/main/course_10

The skill parses the tree URL into repo + --ref + --subdir. A bare monorepo root is refused with a pointer to browse for the course folder; any other bare root proceeds normally.

Framing correction

#2's copy sold this as a way to "skip the manual notebook download". It isn't a replacement — it's an optional supplement that adds the course's real code and helper.py on top of the site download, which stays the main path. Rewritten in new-course/SKILL.md, README.md, and the new skill; /new-course's item 4 handoff is untouched, with the ingest trailing as a clearly subordinate item 5.

Verified

• Backward compatibility — pre-change script vs new, same fixture, no --subdir: byte-identical output
• Scoping — unscoped fixture gives 2 notebooks + 1 cross-course helper.py conflict; --subdir gives 1 + 0
• Sparse clone size against the real monorepo (8.7 MB of 3.4 GB; only course_10 checked out)
• Guards: bad --subdir → exit 1, no partial file; --out into transcripts/ → exit 1
• No clobber — hand-downloaded file's sha unchanged across an ingest
• Tree-URL parse produces the same output as hand-typed flags
• Both refusal paths (cwd guard, monorepo root) fire without invoking git

Not included: the untracked .agents/ Codex mirror was updated locally for parity but is deliberately out of this PR.

🤖 Generated with Claude Code