The multimodal codelab's starter is its own app, not a copy of another codelab's

[DenisovAV]

codelabs/multimodal-flutter-gemma/step_01_starter/ was a byte-identical copy of Getting Started's complete/, enforced by a MIRRORS row. Two things followed from that, and both are worth fixing.

The learner downloaded half a gigabyte and then abandoned it

Getting Started ends on Gemma 3 1B — 0.5 GB, behind a Hugging Face licence gate, so Step 1 needed a token and --dart-define=HF_TOKEN=hf_.... Step 2 then downloaded SmolVLM2 500M instead and never opened the first model again. Accept a licence, wait out 0.5 GB, throw it away.

Worse: the code contradicted the step it was under

Step 1 is titled "A modality is a session flag". But Step 2 swapped the model and set supportImage, so a learner had no way to tell which of the two made the picture work. The codelab asserted its thesis and then demonstrated something else.

The change

step_01_starter is now Step 2 minus the vision flag: the same SmolVLM2 weights, opened as a text session.

model.dart, main.dart and download_page.dart are byte-identical to step_02_vision's. chat_page.dart is that file minus the vision machinery — image_picker, supportImage on both getActiveModel and createChat, Message.withImages, and the token budget, which stays at 1024 here so Step 2's "the call above it changes too — in two ways" stays true.

So Step 2's diff is now the lesson and nothing else. The file on disk does not change between Step 1 and Step 2, which means "a modality is a session flag" is something the learner watches happen rather than something they take on trust.

It also removes the last Hugging Face plumbing from this codelab. The comment in step_02 / step_03 / complete's main.dart still told the reader Step 1 was gated; it isn't, and there is now no token and no --dart-define in any step.

Text

The smolVlm2 block, the ModelType.general paragraph and the main.dart try passage move from Step 2 into Step 1, where the code now lives. Step 2 gains a short section saying what did not change, and why that is checkable only because Step 1 was already running these weights.

Durations move with the material (Step 1 4 → 7, Step 2 12 → 9). The 46-minute total the catalogue carries is unchanged.

Gate

The getting-started/complete → multimodal/step_01_starter row leaves MIRRORS. One row remains — Inference Engines genuinely does start from Getting Started's finished app — so the fail-closed #MIRRORS -eq 0 guard still has something to guard.

Removing it would have traded one unenforced invariant for another, because Step 2's text now tells the learner to diff the two apps and says which files come back identical. MIRRORS cannot express that: it compares a lib/ directory whole, and here exactly one file has to differ — it is the step. A comment added to the starter's main.dart would have made the text quietly false with every other check still green.

So the gate gains SHARED_STEPS, for consecutive steps inside one codelab. Each row names the files that must match and the one that must not, and both halves are asserted: if chat_page.dart ever matched too, Step 2 would have no diff at all, which is as wrong as drift. Proved to fail closed three ways — a comment appended to the starter's main.dart, the starter's chat_page.dart replaced with Step 2's, and model.dart renamed away — each naming the offender, each reverted and re-run clean.

Review round

Three reviewers plus Codex read this branch. Six findings, every one confirmed against the tree before it was fixed.

Two were fail-opens in the SHARED_STEPS check this PR adds — both mine, both the exact class the check exists to prevent. diff exits 0 identical, 1 different, and >1 when it could not compare at all; elif diff -q …; then reads every non-zero alike, so a file that exists but cannot be read took the "they differ" branch and the must-differ half of the invariant passed without being checked. Status 1 is now named in a case, with || status=$? so set -e does not abort before it runs. The other: a half-emptied row ran the shared-file loop zero times and reported nothing. Six perturbations now tested, each red, each reverted.

The ModelType.general paragraph was wrong in both halves, and this PR had promoted it into Step 1. For a .litertlm file extensions.dart:74-77 returns raw on every platform this codelab targets except iOS, and raw never consults modelType — so naming the wrong family does not "wrap the prompt in another model's turn markers", it does nothing at all. On iOS, the one manual target, general and gemmaIt both emit Gemma's own <start_of_turn> markers, so general is not "the template inside the file" either. What the field does decide is post-processing: cleanResponse strips thinking tags for a fixed list of families that includes gemmaIt and not general. The text says that now.

The other three were claims this PR itself made checkable by telling the learner to diff the two apps, and which then failed on sight: "one line changes in _send" (five places do), "the one thing you edit" (contradicted by Step 2's own "in two ways" a page later), and a diff promise that omitted the five plugin registrants flutter pub add regenerates.