Changelog¶

All notable changes to marimo-grader will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Two things release from this repository on their own tags. The service — the backend, worker and web app, shipped as container images — releases on v*. The notebook widget, marimo-grader-client on PyPI, releases on client-v* and is versioned separately, because a student's notebook and the server it submits to upgrade at different times by design. Entries below say which.

[Unreleased]¶

Added¶

Rubrics. A question graded by hand can carry a rubric: Add a rubric on the assignment page takes an outline (parts at the margin, items indented under them, 1 point each unless an item ends in "(2 points)"), pasted from a document if you like, and shows the items it makes before you save. In the grading queue each item gets its own score, grouped under its part with a subtotal, and Full marks for the blank items lets you start from full marks and take points off. The whole-notebook question is worth what its rubric adds up to. Only instructors and TAs see a rubric; scores outside an item's range, or for an item the rubric does not have, are refused.

Fixed¶

Client 0.2.4: put() stores figures, brain maps and more tables. A .nii.gz name is now gzipped: nibabel's to_bytes() is uncompressed, so an image stored that way could not be read back. nltools BrainData is accepted (to_nifti()). .png, .jpg, .svg and .pdf names take a matplotlib figure or axes, a seaborn grid or a nilearn plot. .csv/.tsv take polars DataFrames, nltools DesignMatrix and numpy arrays as well as pandas, whose row labels are kept unless they are the default 0, 1, 2…. A str under a binary name (a file name passed by mistake) is refused with a pointer to upload() instead of being stored as the file.

[0.1.10] — 2026-10-19¶

Changed¶

The assignments table shows where each assignment stands. Rows are grouped into Open now, Coming up and Closed (folded); one When column gives the date that matters next; a bar shows the class's progress (graded, waiting for you, failed, not submitted) and an Average column the mean score of students whose work is fully graded. Only the unusual gets a label: Hidden from students, Enrolled only, Not re-run. Each row has Grade (with the count waiting, straight to the queue when one question has them) or Review, Settings and More (copy student link, preview as a student, upload a corrected notebook, export to Canvas). Filters and a search sit above it. Rows are selected by clicking them (shift-click for a run); a bar at the bottom exports the selection or changes its due date.
The assignment page is down to what is used. A sentence on where it stands, then one row per question with a strip of marks, one per student, colored by state; a mark opens that student's answer in the queue (/grade/<question>?netid=). The settings summary, student links and version history are gone from the page: Settings is a side panel, shared with the table, and the rest is under More.
Settings are grouped into When, Who and where, and Attempts and grading, with the title editable alongside.

Added¶

GET /offerings/{id}/progress[?assignment_id=]: each student's state on each question (graded, waiting, failed, running, none), counted per question and per assignment, with the answers to grade and the average; with an assignment, the per-student marks in roster order (services/progress.py).

[0.1.9] — 2026-10-19¶

Changed¶

Publish a new version from the assignment page, with the instructor notebook alone. The grader builds the student copy and reads the questions as grader publish does (services/student_copy.py, now shared with the CLI). A preview comes first -- the version number, the questions and how each is graded, what is new, removed or changed, the student copy to download -- and Publish version N makes it live (POST .../versions/preview, POST .../versions/from-notebook). The hand upload of two notebooks plus question JSON is gone from the page; its API remains for the CLI.

Fixed¶

Version history times carry their UTC offset; they showed as hours in the future on servers whose database returns naive timestamps.

[0.1.8] — 2026-10-19¶

Changed¶

Assignments is a table. One row per assignment with its access, status, opens / due / closes, version, questions, points and grading; click a column to sort, again to reverse (due date by default). Click rows to select them (or Select all) and Export Canvas CSV takes the selection. The title opens a new assignment page: key settings, a Settings button that opens the form in place, the student notebook, Open in molab and a copyable launch link, the questions with their grading queues, and the version history (GET /offerings/{id}/assignments/{aid}/versions). Uploading a version by hand moved there, under the grader publish command that is the usual way.

[0.1.7] — 2026-10-19¶

Changed¶

Needs attention groups grading failures by cause. One entry per assignment and cause instead of one line per submission, saying what happened, whose move it is, and whether a retry can help (services/failures.py). Retry all re-queues a group (POST /offerings/{id}/retry), and the page follows each retried submission to its outcome -- queued, grading, graded, waiting for a hand grade, or failed again with the new reason (GET /offerings/{id}/retry-status). The awaiting-grade line no longer repeats a question id used as its title.

[0.1.6] — 2026-10-19¶

Security¶

The safety scan excuses only findings the release also had. 0.1.5 excused a flagged line whenever its text appeared anywhere in the released copy, so a line the release merely contained (in a string, say) could cover a student's import. A finding is now excused only when the released copy was flagged for the same thing on an identical line.

[0.1.5] — 2026-10-19¶

Changed¶

Renders show the code. The worker renders submissions with --include-code, so the grading view (and a student's own) shows each cell's code above its output instead of only "passed". The submission view has Rendered / Code tabs, the second being the source as submitted. grader rerender re-renders older submissions one at a time while the queue is otherwise idle.

Fixed¶

A student is not refused for a line the course gave them. The safety scan now runs before grading and excuses a flagged line that appears, as written, in the released copy. A thresholding release shipped import io in its setup cell, and every submission from it failed before running.

[0.1.4] — 2026-10-19¶

Added¶

Student preview. Instructors and TAs can switch their own view of an offering to a student's (PUT /offerings/{id}/preview, Preview as a student on the Student view page) and test an assignment the way the class will: the student list and portal, close_at and the deadline enforced, student-scoped storage credentials, staff pages refused -- in the web app and in notebooks signed in as them. Release dates do not apply, so an exam can be tried before it opens; submissions remain test attempts. A banner offers Exit preview. New enrollments.preview_student column (migration c4d8e1f2a903).

[0.1.3] — 2026-10-19¶

Added¶

Private assignments and exams. visibility: enrolled keeps an assignment out of assignments.json and serves its notebook (/a/.../student.py, /molab) only to signed-in members of the offering, by session or notebook token. release_at / due_at / close_at form one window for the notebook, its storage data and submissions; staff are exempt. A private assignment without a release_at stays closed and is left out of students' assignment list, so an exam can be built in place. grader publish sends settings with the create, so a new private assignment is never briefly public. late_policy: reject refuses work after the deadline, otherwise it is accepted and marked late. A new /a/<course>/<term>/<slug>/launch link signs a student in and lands on their portal, which now offers Open in molab and Download .py and shows the window. Storage credentials and presigned URLs no longer outlive an assignment's close_at. Docs: For instructors → Exams and private assignments.
Whole-notebook submission. notebook_submission adds a notebook question that takes the entire notebook: manual when notebook_points > 0, a graded-on-arrival receipt at 0. Turning it on lets a version publish without a _marks cell; notebook is a reserved qid.
autograde: false stores and renders submissions without running checks, and render_timeout gives an assignment longer than the worker default. render: false skips the re-run too: the grading view shows the submitted code and a Saved files panel of what the student wrote under storage.private()/<slug>/ (GET /submissions/{id}/files, presigned links, files changed after due_at flagged).
grader publish gains --private/--public, --release, --due, --close, --late-policy, --no-autograde, --notebook-points and --render-timeout.
Client 0.2.0: course storage in the notebook. marimo_grader_client.storage is the notebook side of the storage broker, moved here from dartbrains-tools so any course on the grader gets it: signin_button() / connect(), course(), private(), group(), assignment(slug), a durable @cache, local_path() for anything that wants a file. One obstore store per mount with a self-renewing credential; the token is cached where the runtime keeps things (~/.config/marimo-grader, or molab's .env, which it recognises by its sandbox). GRADER_* names throughout; storage.configure(server=...) for a course's default. assignment_card(slug) links a chapter to its assignment on the grader. dartbrains-tools 0.3 wraps this and keeps its imports working.

Fixed¶

Client 0.1.2: a sign-in token restored from the browser's storage now reaches the kernel as well. The widget showed "signed in" but left the token trait empty, so Python code keyed off it -- the storage layer's connect() -- saw nobody. Only a fresh click synced it.

[0.1.2] — 2026-09-18¶

Added¶

Storage broker (service). A signed-in notebook can ask the grader for storage credentials: POST /offerings/{id}/storage/session returns two Cloudflare R2 temporary credentials scoped to exactly the prefixes the enrollment may reach — course data, released assignments and the shared cache read-only; the student's own area, group folders and private cache read-write; staff additionally the student areas they may grade. POST …/storage/presign gives a single presigned URL for runtimes that cannot sign requests, GET …/storage/me the mount table without minting. The grader stays in the control path only; bytes flow between the notebook and R2 directly. Student prefixes are an HMAC of the NetID, so a listing never exposes the roster. Assignments gain optional release_at and close_at settings; outside that window the prefix is absent from every credential, so pre-release data is unreachable rather than hidden. Off until GRADER_STORAGE_ENABLED and the Cloudflare values are set; see the configuration reference and docs/storage-architecture.md.
New tables groups, group_members and the append-only storage_grants audit (migration b7c2e9d41f03).
scripts/r2_spike.py, the isolation experiment the design rests on.

[0.1.1] — 2026-09-17¶

Security¶

Every login-link code is now refused at the device poll, not just the operator one. poll() compared the code's client against the literal "operator-login-link" rather than checking LOGIN_LINK_CLIENTS, so only half the set was excluded. Both kinds of login-link row are minted pre-approved and unconsumed, and the poll mints a notebook token from any such row — so POST /auth/device/token, given the code out of a sign-in email, returned an eight-hour notebook token without the link ever being opened. An email code lives seven days; a device code lives ten minutes. Impact was limited — the holder of the code is the intended user, and the token grants only their own submit-and-read access — but the two paths are meant to be disjoint, and consume_login_code() already checked the set. (#19)

Removed¶

A roster test helper, _netids, that nothing called and that could not have worked if anything had: the endpoint it read returns a list, so its first line would raise. (#18)

Documentation¶

The documentation site was reworked (marimograder.org): a live demo assignment that runs in the reader's browser alongside the instructor notebook it was generated from, an Authentication section covering the device handshake, SAML, email sign-in links and token lifetimes, the deployment runbook moved out of deploy/README.md and onto the site as eight operator pages, and a How it works section. (#14, #15)
Corrections found by reviewing the new pages against the code: an unreleased score is provisional rather than hidden — the student portal shows it and the widget labels it — grader seed defaults --students to two example NetIDs, warm-cache scans the version's instructor artifact rather than the published copy, the Ed25519 key-type error is raised at first token mint rather than at startup, and GRADER_USE_BUBBLEWRAP=1 enables the sandbox without requiring it. (#15)
The sidebar labels the index page "Home" instead of repeating the site name, and the header carries marimo-book's mark until the grader has one of its own. (#16, #17)

[0.1.0] — 2026-09-16¶

First tagged release: the service as deployed for the DartBrains pilot — SAML and notebook sign-in, publishing, immutable per-attempt submissions, sandboxed autograding, the grading web app, roster import and Canvas export.