The study panel

Five sections, in order of certainty

Open a planet and the panel builds itself from the most certain evidence outward:

1. What this is — a summary written from parser facts alone: kind, file and line, size, how many things use it, how many it uses, and how many of those links are possible rather than certain. No key, no network, no model.
2. The explanation — grounded narration from your configured provider, with a file:line citation on every claim. Codemble refuses to display provider output that names anything outside the parsed graph.
3. Connections — every relationship the parser observed into and out of this structure, grouped inbound and outbound. Each row states direction, whether the relationship is certain or only possible, and the file:line where the other structure is defined, so you can go read it. Click any row to study that structure. A small diagram above the list shows callers, this structure, and callees at a glance.
4. Real source — the exact lines, numbered, straight from your file.
5. The language lens — idiom notes anchored to constructs the parser actually detected.

Sections 1, 3, 4 and 5 never involve a model. If narration fails or is not configured, they are all still there.

Easy and Expert

The header’s Mode toggle changes how Codemble talks to you, and how much it puts on screen at once:

	Easy	Expert
Narration	Short sentences, every term explained in place	Concise, assumes fluency
Check questions	“Which piece of code…”	“Which structure…”
Labels	“Called by”, “Possible connection”	“Callers”, “Possible relationship”
Lens notes	The idiom in plain words	The precise language mechanic
Density	Opens on the Map, hides unrelated galaxy edges, larger type	Opens on the Galaxy, shows everything

Some surface names change with the mode as well: the flat layer is labelled Diagram in Easy and Map in Expert, and its two tabs read How it fits together / What runs first in Easy against Architecture / Workflow in Expert. They are the same two views of the same parser evidence; this documentation uses the Expert names throughout.

Easy mode also shows a hint chip naming the nearest unlit region to Home, counted in import-route hops over the graph — no model picks it for you. Where several regions sit the same number of hops away, it prefers the one with more parser-proven structures, so the first suggestion is rarely a one-line package init. Its action follows the current level: open the suggested system, read its source, then prove it. Once the correct next step is already on screen, the chip becomes an instruction instead of leaving a button that cannot move you forward. Guidance stays out of the way until the first-run choices — audience, Home, and the coach marks — are finished.

Mode never changes the graph, the coordinates, your progress, or how a check is scored: both question voices are generated up front and scoring compares option IDs, which have no voice at all. It is remembered per project.

No key? Nothing important is missing

Codemble is bring-your-own-key. Without one, the panel says so and everything except the narration prose keeps working.

To narrate without sending your code anywhere, use a local model:

ollama pull gemma4:12b
export CODEMBLE_PROVIDER=ollama
export CODEMBLE_OLLAMA_MODEL=gemma4:12b

The panel tells you whether Ollama is already running on this machine and which model it recommends. Honest caveat: grounding validation catches an invented identifier, not a wrong claim about a real one, and smaller local models make that second kind of mistake more often.

Partial parses

If a file has a syntax error, Codemble keeps it visible and refuses to invent structure inside it. Narration stays off for that file, and both the structural summary and the narration block say why.