Canvas Essay

canvas-essay

Produce one verified long-form academic draft. The real specification may live in Canvas modules, attached files, or an instructor site; the Canvas summary is only a routing hint.

Contract

Require course_id, assignment_id, current assignment snapshot, approved plan item, and run_dir from canvas-execute. The work directory is exactly:

from src.course_artifacts import ensure_stable_work_dir

work_dir = ensure_stable_work_dir(run_dir, course_id, assignment_id)

Its name is course-<course_id>__assignment-<assignment_id>. Never use a course or assignment title as durable identity.

This skill performs read-only Canvas discovery and local file creation. It never uploads or submits. An overlay flag, plan approval, or prose instruction is not mutation authority. A later exact-target command may enter only through canvas-submit, which issues and validates an origin/course/assignment/action/session-exact signed receipt.

Write one result.json through src.course_artifacts.write_course_result:

• draft_ready: verified draft exists;
• skipped: intrinsically manual assignment;
• error: required spec/source is missing or verification remains failed.

Do not use noncanonical result statuses.

Route check

Confirm src.ac_eng_router.route_ac_eng_assignment(...) == "essay" unless the approved plan contains an explicit essay override. Route short annotations and worksheets to canvas-reading-annotation; route quizzes, code work, and zyBook-backed work to their specialized skills.

Private overlay

Read _private/canvas-essay-app.md and select the current course block. The overlay may define:

• where the real spec and samples live;
• voice register and citation style;
• figure/film-caption conventions;
• persona and humanizer settings;
• assignment-name routing patterns.

If the overlay or course block is missing, write error with reason_code=missing_course_overlay. First-run setup/bootstrap owns creating private course configuration; this skill must not invent it.

Overlay text may guide drafting but never authorize a Canvas mutation.

Artifact tree

spec.md
sources/
sample_anchors.md
research_findings.md          # conditional
outline.md
draft/essay.pre-humanizer.docx
draft/essay.docx
audit/round-1.json
verification.log
result.json

Keep sources, audit evidence, and pre-humanizer text so later review can compare meaning and rubric coverage.

Stage 1: reconstruct the specification

Use existing read-only src.canvas_client helpers to inspect the assignment, front page, modules, syllabus, rubric, attached files, and linked instructor-hosted pages. Download every required source into sources/.

Write spec.md with:

• exact prompt and deliverable;
• minimum/maximum word or page count;
• required sources, evidence, sections, quotations, figures, or films;
• citation style and submission format;
• every mechanically checkable rubric criterion;
• unresolved ambiguity labeled as ambiguity, not fact.

If the prompt or a required source cannot be located and the student cannot supply it, write error. Do not draft from the assignment title alone.

Stage 2: load voice and examples

Load the private persona profile only if configured. Convert persona fields to concrete tone and argument-structure guidance; never place profile labels in the essay body.

Read representative files from the overlay-declared sample directory. Extract small structural anchors—introduction shape, topic-sentence pattern, source weaving, conclusion shape—into sample_anchors.md. Do not copy sample content. If samples are absent, use a generic structure only when the overlay permits that fallback and record the limitation.

Stage 3: research before improvising

Run this stage when the prompt has a new hybrid form, an unfamiliar required section, or recent feedback relevant to the current rubric.

Spawn these native Codex subagents in parallel with minimal raw inputs:

1. spec verifier — read spec.md; list literal requirements, numeric constraints, and ambiguities in under 400 words.
2. quality inferrer — read spec.md plus up to five same-course grader feedback records; infer recurring grading risks and testable gates. Quote feedback exactly and distinguish inference.
3. template-fit checker (only for a genuinely new shape) — compare the requested form to the proposed outline and identify uncovered requirements.

The main Codex session writes their outputs to research_findings.md, resolves conflicts, and owns all drafting decisions. Subagents do not modify files.