check-cell-order

---

name: check-cell-order description: Detect cell-ordering / enchainement problems in Jupyter notebooks (canonical-order slippage, misplaced or forgotten cells). Arguments: [target] [--severity HIGH|MED|LOW] [--json]

Check Cell Order

Scan notebooks for cell-ordering and enchainement (flow) problems — the class friction reported after the SymbolicLearning session: "certaines choses n'étaient pas à leur place" (canonical-order slippage and forgotten / misplaced cells that slip through upstream review).

Backed by scripts/notebook_tools/scan_cell_ordering.py, which reads the beginning and the end of every cell and reports graded findings, each citing the offending text so it is spot-checkable. Epic #3240.

Target: $ARGUMENTS

Arguments

• target: a notebook path, a family subpath (--family SymbolicAI/SymbolicLearning), or --all
• --severity HIGH|MED|LOW: only show findings at or above this level
• --json: machine-readable output (for piping / CI)
• --fail-on HIGH|MED|LOW: exit 1 if any finding at/above this level (default HIGH)

What it detects

Severity	Category	Meaning
HIGH	SECTION_ORDER	numbered markdown headers go backwards under the same parent (## 3. then ## 2.)
HIGH	EXERCISE_ORDER	Exercice N / Exemple N labels out of order
MED	DANGLING_INTRO	a markdown cell ends announcing imminent code, but the next cell is not code (forgotten / misplaced cell)
MED	INTERP_BEFORE_CODE	an interpretation markdown sits before the code it comments instead of after its output
LOW	SECTION_GAP	a numbered header skips a value (possible omission)
LOW	CONSECUTIVE_CODE	more than 3 code cells in a row with no markdown between

Legitimate numbering is silent by design: a normal increment (1→2), a sub-section open/close (3→3.1→4), a reset-to-1 group restart, a level-mixed H1 title (# 13. series number vs ## 1. sections), fenced-code comments, and TOC / overview cells listing ≥2 labels.

Process

1. Run the scanner on the target:

   python scripts/notebook_tools/scan_cell_ordering.py <target> --severity HIGH
   # or a whole family:
   python scripts/notebook_tools/scan_cell_ordering.py --family SymbolicAI/SymbolicLearning
   # or the whole repo:
   python scripts/notebook_tools/scan_cell_ordering.py --all --severity HIGH
   

   On this machine the script needs a Python with the stdlib only; any of the notebook kernels works (e.g. C:/Users/MYIA/AppData/Local/Programs/Python/Python310/python.exe).

2. Ground-truth every HIGH finding before acting (rule G.1 — the finding is a signal, not a verdict). Dump the actual headers / labels in document order to confirm it is a true positive, e.g.:

   python - <<'PY'
   import json, re
   from pathlib import Path
   H = re.compile(r"^(#{1,6})\s+(\d+(?:\.\d+)*)\b")
   nb = json.loads(Path("<notebook>").read_text(encoding="utf-8"))
   for i, c in enumerate(nb["cells"]):
       if c.get("cell_type") != "markdown":
           continue
       for ln in "".join(c["source"]).splitlines():
           if H.match(ln.strip()):
               print(f"cell#{i}", ln.strip()[:70])
   PY
   

3. Fix or justify:

   • True positive → re-order the cells with NotebookEdit (work bottom→top to avoid index shift). Re-order clears outputs, so re-execute the affected notebook (Papermill, rule C.2) before commit.
   • Defensible by design (rare residual FP) → leave it; if recurrent, harden the scanner and add a regression case to scripts/notebook_tools/tests/test_scan_cell_ordering.py.

4. Re-scan the notebook / family to confirm 0 unjustified HIGH before commit.

Conventions

• Canonical ordering convention: see .claude/rules/notebook-conventions.md ("Structure pédagogique" + "Enchaînement des cellules").
• Classification of Exercice vs Exemple is content-based — do NOT relabel to silence a finding (see .claude/rules/exercise-example-labeling.md).
• C.1/C.2/C.3 still apply: no raise NotImplementedError in stubs; commit with outputs; only stage notebooks whose source changed.

CI

The scanner is wired into a per-PR gate that scans the notebooks changed by a PR and fails on HIGH (--fail-on HIGH). Pre-existing findings in untouched notebooks are tracked separately under Epic #3240 and do not block unrelated PRs.

Notes

• The scanner is read-only — it never edits notebooks. Fixes are a separate, deliberate step.
• Exit codes: 0 clean / 1 finding at/above --fail-on / 2 usage or IO error.
• A malformed notebook prints ERROR <path>: <reason> and the scan continues.