CodeRunner and Jobe
CodeHelp can be used alongside Moodle CodeRunner in two complementary ways:
- CodeRunner remains the grading and assessment tool.
- CodeHelp explains visible feedback and guides the student’s next debugging step without writing a corrected solution.
The simplest setup is to add a CodeHelp LTI link next to each CodeRunner activity. Instructors can create a context for the activity and use the context’s share link so students arrive at the help form with the right assignment context already selected.
Visible CodeRunner Feedback
Students can paste visible CodeRunner feedback into the CodeRunner/Test output field on the help form. This is sent to the tutoring prompt as execution evidence. Students should not paste hidden tests, model answers, or private instructor-only feedback.
Good feedback to paste includes:
- compiler messages
- runtime errors
- stdout/stderr from a visible test
- public test-case feedback shown by CodeRunner
Optional Jobe Execution
If JOBE_BASE_URL is configured, Nur ausführen calls /help/run-code for supported programming languages without invoking an LLM or storing a help query. Standalone runs are recorded server-side with an opaque run ID, bounded result and a fingerprint of the exact execution inputs; raw source and stdin are not additionally stored in this evidence record. When execution is selected for a help request, CodeHelp reuses a matching, unexpired server result or performs a fresh Jobe run and stores that result with the query. Reuse defaults to 24 hours and requires the same user, course, role, task, context, version, language, source, inputs and public setup. Changed, expired or transport-error evidence triggers a fresh run; browser-supplied output alone is never trusted. Expired records can be cleaned up, and user/course deletion removes associated run records. The LLM receives observed run evidence, such as language, outcome, stdout, stderr, and compiler output.
Keep the Jobe server separate from the Gen-Ed application server and firewall it appropriately. If the Jobe server requires API-key access, set JOBE_API_KEY.
SQL exercises
The language selector includes SQL (SQLite). This uses a fixed Python 3.11+ runner with the standard sqlite3 module on Jobe; it does not require a native Jobe SQL language. The instructor’s context supplies public SQLite setup SQL separately from the human-readable schema and sample data. Each run starts with a new in-memory database. One learner statement is accepted per run, and result size and computation are bounded. Attaching files, PRAGMAs, and extension loading are blocked.
Nur ausführen calls /help/run-sql without invoking an LLM or storing a help query. Authentication, active course, context release and context version are checked before using the server-side setup. Client-provided setup is ignored. Selecting execution when submitting a help query reuses only matching fresh server evidence; otherwise it performs a new run. The output is stored with the tutoring query. The UI labels SQLite explicitly even when the teaching dialect is MariaDB/MySQL. It never treats execution success as a grade.
Guided learning and teacher drafts
The local workspace offers a teacher-configured public example and five freely selectable phases: predict, run, investigate, modify and make. The first three show read-only example code; the last two allow editing. Optional AI help and a separate free-help mode remain available. Each mode retains its own draft while the page stays open. Help queries store three bounded reflection fields, the selected phase and an optional strictly scoped parent query reference.
The context editor’s AI authoring tool follows three steps: Auftrag → Prüfen → Speichern. It defaults to a complete task package, with individual areas available as optional scopes. Teachers edit draft fields, apply selected fields and undo the last application. Private outputs start unselected; existing private notes and solutions are excluded from generation inputs. Generation uses the course model and usage ledger. Generating or applying a draft never saves, releases or executes it automatically.
During review, teachers can explicitly test either the public example or the internal reference solution through Jobe without saving or releasing the context. Up to five cases accept stdin and an optional expected stdout. Comparison normalizes line endings while preserving other whitespace and case. SQL creates a fresh database from the supplied public setup for every case and accepts no stdin. Transport failure, execution failure, truncated output and successful output comparison remain distinct. Editing relevant inputs invalidates displayed test results. These tests do not save the draft or count as learner attempts; they do not establish teaching suitability or correctness beyond the supplied checks.
Existing installations need all pending migrations, including 20261115--codehelp-learning-cycle.sql and 20261115--codehelp-run-evidence.sql. As of 15 November 2026, deployment of the combined implementation is in progress; successful production deployment and verification are not yet confirmed. See Unterrichtsprofile und Lernbegleitung for setup.
Instructor Cockpit
Instructors can open the CodeHelp Instructor-Cockpit from the help form. It aggregates:
- number of CodeHelp queries
- number of students using the tool
- queries containing visible CodeRunner feedback
- Jobe runs stored with help queries and their outcomes
- frequent extracted difficulty topics
- contexts with the most help-seeking activity
The cockpit also shows help requests without a linked follow-up attempt. Standalone runs now have separate server-side evidence records; query-linked run counts retain their stated scope.
The explicit KI-Briefing erstellen action asks the course model for two to three teaching actions and two discussion questions. It uses aggregate counts and fixed anonymized topic, hint-level and feedback labels from at most twelve matching recent query records. Free issue/hint text is classified locally, never sent verbatim. No learner names, raw code, private teacher context or model solutions are included. Every suggestion cites an authorized query ID; malformed output is rejected. Empty selections do not invoke a model. The proposals are not diagnoses or grades.
Technical checks and synthetic evaluation do not establish learning effectiveness. The German classroom evaluation guide describes public LF05, LF22 and SQL cases, teacher rubrics and explicitly bounded model runs. A real classroom validation has not yet been conducted. An unlinked follow-up is not evidence of abandoning the task.