coldcall_lti/README.md

# coldcall-lti

A Canvas external tool (LTI 1.3) for managing cold calls in case-based
classes: it selects students to call using weighted randomness, records
what happened with each call, lets students report planned absences, and
reports participation data back to both instructor and students. It
replaces a manual workflow built on exported rosters, Google Forms, and
local scripts.

## How it works

The tool is a Flask application that Canvas launches over LTI 1.3. A
single URL serves everyone: Canvas identifies the person and course on
each launch, so instructors get the call-list and reporting views while
students get the absence form and their own history. The roster comes
from Canvas through the Names and Role Provisioning Service, which means
adds and drops are picked up automatically rather than reconciled by
hand.

Selection uses the same weighting as the manual system it replaces: each
answered call divides a student's weight by the course's weight factor
(default 2), so students who have answered more questions become
progressively less likely to be called. Each course can instead use
"cycle" mode, which shuffles the roster and calls everyone exactly once.
These, along with whether students can see their own assessments, are
per-course settings.

## Layout

- `coldcall_lti/models.py` — SQLAlchemy models: courses (with their
  settings), students, enrollments, opt-outs, and calls, plus the query
  helpers that feed selection.
- `coldcall_lti/selection.py` — the weighted and cycle selection logic.
  Kept free of database and web dependencies so it can be tested and
  reasoned about on its own.
- `coldcall_lti/__init__.py` — the Flask application factory.
- `migrations/` — alembic migrations. The schema avoids
  database-specific types so the same migrations run on SQLite (the
  default) and MariaDB/MySQL; switching is a matter of changing
  `COLDCALL_DATABASE_URL`.
- `tests/` — pytest suite covering selection behavior and the model
  helpers.

## Setup

Development uses a virtualenv that shares the system's Debian-packaged
libraries (Flask, SQLAlchemy, alembic, pytest) and adds the one
PyPI-only dependency, the maintained `pylti1p3next` fork of PyLTI1p3:

```
python3 -m venv --system-site-packages .venv
.venv/bin/pip install pylti1p3next
.venv/bin/pip install -e .
```

Create the database and run the tests:

```
.venv/bin/python -m alembic upgrade head
.venv/bin/python -m pytest tests/
```

Configuration is by environment variable: `COLDCALL_DATABASE_URL` (any
SQLAlchemy URL; defaults to an SQLite file under `instance/`),
`COLDCALL_SECRET_KEY` for Flask sessions, `COLDCALL_LTI_CONFIG` (path
to the LTI platform configuration), and `COLDCALL_DEV_MODE=1` to enable
the fake-launch pages.

## Connecting to Canvas

The tool speaks LTI 1.3, which requires a Developer Key created by a
Canvas account admin. The key points Canvas at three endpoints here:
`/lti/login` (OIDC initiation), `/lti/launch` (the launch target), and
`/lti/jwks` (this tool's public keys). The platform side is described
in a JSON file — copy `lti_config.example.json` to
`instance/lti_config.json` and fill in the client id and deployment id
from the Developer Key. Generate the tool's keypair alongside it:

```
openssl genrsa -out instance/private.key 4096
openssl rsa -in instance/private.key -pubout -out instance/public.key
```

On each instructor launch the tool refreshes the course roster from
Canvas through the Names and Role Provisioning Service, so enrollment
changes appear without any manual step. There is also a "Sync roster
now" button on the instructor page, and a management command suitable
for an hourly cron job on the server, which keeps rosters current even
when nobody has launched the tool (worth having during the add/drop
churn at the start of a term):

```
17 * * * * cd /path/to/coldcall_lti && .venv/bin/flask --app coldcall_lti sync-rosters
```

Courses whose Canvas end date has passed are skipped automatically.

Student names come from Canvas display names, which already reflect
preferred names. Pronouns require one extra piece of Developer Key
configuration: add a custom parameter

```
pronouns=$com.instructure.Person.pronouns
course_start=$Canvas.course.startAt
course_end=$Canvas.course.endAt
grading_scheme=$com.instructure.Course.gradingScheme
```

The course dates bound the schedule and date pickers; the tool works
fine without them when a course has no dates set in Canvas.

to the key. Canvas then includes each person's pronouns in launches
and in the roster data (the tool requests memberships scoped to the
resource link, which is what makes Canvas attach per-member custom
fields). Pronouns appear on the live call card, printed call lists,
and each student's own page. If the account has pronouns disabled,
everything simply shows without them.

## Opt-outs

Students remove themselves from a day's cold-call list by picking the
date on their page; nothing else is asked. There is deliberately no
enforced submission deadline: an instructor running live calls in
class gets up-to-the-second opt-outs automatically, while one who
prints a call list beforehand should just tell students how much lead
time they need (for example, "an hour or two before class"), since
opt-outs after printing won't be on the paper.

Withdrawing an opt-out is bounded, because un-opting-out after the
fact would rewrite a student's availability record (and with it their
participation grade): withdrawal closes when class begins, using the
start time recorded on the schedule page, or at the start of the class
day when no time is recorded. Withdrawals are also recorded rather
than deleted — the opt-out export includes a withdrawn_at column, so
the full history of changes survives.

## Grading

Final participation grades use the timing-neutral foregone-
participation scheme (a port of the earlier participation_grades.R):
a student's grade is the quality of their answers minus a deduction
for participation they missed by being unavailable, estimated by
Monte Carlo simulation of the actual weighted draw and averaged over
when the absences fall, plus a small incentive penalty for being
drawn while absent with no opt-out filed. The reason for an absence
never matters, and luck of the draw never moves a grade. Parameters
(allowance, passing line, penalties, simulation size) are per-course
settings; grades are computed on demand, reviewed on the instructor's
grades page, and shown to students only when the instructor publishes
reports. With gradebook passback enabled, a review-then-push page
sends the reviewed scores to Canvas via the Assignment and Grade
Services; nothing is ever sent without explicit confirmation.

Grades are computed in points out of 100 and displayed through a
per-course scale: the built-in linear UW 4.0 map, a threshold table
(letter grades, importable in one click from the course's own Canvas
grading scheme), or raw points.

The port was verified against the R engine's rendered reports from a
real course: answer quality and availability match exactly; the
simulated penalty agrees within Monte Carlo noise. `flask --app
coldcall_lti import-legacy <dir>` imports a manual-era class directory
for this kind of testing.

## Developing without Canvas

Because a Developer Key takes institutional approval to get, the app
has a fake-launch mode for local development:

```
COLDCALL_DEV_MODE=1 .venv/bin/flask --app coldcall_lti run --debug
```

Then open http://localhost:5000/dev and launch as the fake instructor
or any of the fake students. This sets up exactly the session state a
real launch would, and the fake roster flows through the same sync code
as real NRPS data, so everything past the launch behaves identically.
Dev mode also relaxes the cookie settings that Canvas's iframe
embedding requires in production (SameSite=None; Secure), which would
otherwise break plain-http localhost use.