Canvas Skill Opportunity

Canvas Skill Opportunity

Choose the best first durable workflow, not merely the easiest-looking assignment. This is an Agent judgment protocol. Deterministic code may discover recurrence and project private data, but it must not decide the recommendation.

Contract

• Enter after canvas-setup verifies authentication and finds no ready routes, or when the user explicitly asks for skill opportunities.
• Write only:
  • runs/<today>/skill-opportunities.json
  • runs/<today>/skill-opportunities.md
• Keep .claude/ read-only.
• Never start or consume a quiz attempt merely to classify it.
• Never solve, draft, answer, upload, submit, create a route, or create a skill.
• Never call canvas-scan, canvas-execute, or a course-solving skill.
• Stop for one numbered user choice. Selection authorizes only the later canvas-bootstrap design check; it does not authorize assignment submission.

1. Establish The Read-Only Boundary

Read AGENTS.md, courses.yaml, and setup state. Verify Canvas authentication if it has not been verified in this session. Before inspecting course data, say:

I will read recurring assignment instructions and Canvas feedback settings to
recommend one reusable workflow. I will not do work, start an attempt, upload,
submit, or create a skill.

Before writing, verify runs/ is gitignored. Stop if it is not.

2. Discover Recurrence As Fact

Use the existing factual helpers:

from src import canvas_client as cv
from src.opportunity_evidence import derive_quiz_feedback_capabilities
from src.recurring_patterns import bucket_recurring, is_course_active

For each active course, call cv.list_assignments_for_opportunity(course_id) once, apply the existing 7-day term grace period, and call bucket_recurring(items, min_freq=3). The dedicated list helper excludes student submission state. Retain the assignment IDs belonging to each returned bucket, future occurrence count, known future points, and current route coverage. These are evidence and later tie-breakers, not a verdict.

Do not infer task family, response length, or suitability from course name, assignment title, or submission_types alone. A Canvas upload may be code, a Word accounting worksheet, or a long essay; discovery cannot tell which.

3. Inspect Representative Real Specifications

For each plausible recurring bucket, use cv.get_assignment_spec_for_opportunity(course_id, assignment_id) to read enough representative assignments to establish whether the response pattern repeats. This helper returns an allowlisted real-spec projection and does not request a submission include:

1. Read the full assignment description and rubric.
2. Inspect attachment metadata and read required attached Word, Excel, PDF, starter-code, data, or template files when accessible.
3. Follow relevant module pages, Canvas pages, and instructor/source pointers to the real specification. Do not treat a thin Canvas description as the specification.
4. Compare at least two instances when two exist. Read additional instances when their required deliverables disagree. If stability still cannot be established, use insufficient_evidence.

Do not replace this helper with generic get_assignment, which requests an embedded submission object. If another source unexpectedly embeds student state, project it away before the Agent inspects or stores the specification.

Record derived facts, IDs, and source locations in the private report; do not copy full bodies, rubric text, or attachment contents into it. Never read or retain a student's prior answer content for task classification.

Extract these facts:

• actual task family and primary deliverable
• stable response pattern and whether the required inputs are reachable
• Canvas-native or file-upload delivery path
• number and independence of response units
• estimated length and role of the main continuous prose unit
• required live, physical, group, proctored, or unsupported external work
• plausible pre-submit checks
• post-submit feedback and retry policy, with evidence confidence

4. Apply Broad Task-Fit Priors

Treat the following as strong first-skill candidates when their response pattern repeats and their inputs and Canvas delivery path are reachable:

• Canvas-submittable code of any length, even without supplied tests
• mostly objective Canvas quizzes: choice, true/false, matching, numeric, and other objectively checkable items
• quantitative math, accounting, economics, statistics, finance, and business work; a formal regression verifier is not required
• structured Word, Excel, or PDF accounting/business work, worksheets, tables, calculations, and template-driven documents
• independent short answers and short reading/writing annotations

Use response shape, not aggregate word count. Twenty independent annotations of 10-20 words remain a strong candidate even though their combined total may exceed 200 words. Conversely, a main or central continuous prose unit around 200 words or more is a strong default demoter for the first skill because voice, coherence, and review cost compound. This is not a mechanical universal hard gate: ancillary prose in a code, accounting, or quantitative deliverable must not make the central task look like an essay.

Long essays, research prose, personal reflection, and creative writing usually belong in assist_only. Required external-site interaction, live performance, physical work, proctoring, or group participation is unsupported unless an already-verified Canvas Pilot path can complete the whole required deliverable.

5. Evaluate Review In Two Separate Layers

Pre-submit review

Identify reasonable checks without demanding an external validator:

• code: run, compile, inspect interfaces, examples, and constraints
• accounting/finance: recompute, reconcile, balance, and cross-foot
• math/statistics/economics: independent derivation, units, assumptions, direction, and magnitude checks
• structured documents: required-field, source, calculation, rubric, and format coverage
• objective questions: independent solving and source cross-checking

Formal tests improve confidence but are not an eligibility gate. Long code without supplied tests stays strong when its specification, inputs, and Canvas submission path are complete.

Post-submit feedback and retry

For quizzes and other retryable work, separately establish:

• allowed_attempts
• whether results appear before the next attempt
• total-score visibility
• item-correctness visibility
• correct-answer visibility
• own-answer visibility
• feedback timing/window
• scoring policy (keep_highest, latest, average, or unknown)
• question reuse versus randomization between attempts
• evidence confidence and source

Use derive_quiz_feedback_capabilities(quiz) on static Canvas settings first; it is a pure interpreter and never calls Canvas or starts an attempt. If an already-completed sibling assignment can provide observed evidence, call only cv.get_submission_feedback_observation_for_opportunity(course_id, assignment_id). That dedicated wrapper may inspect the existing submission internally, but it returns only a minimal Boolean/enum projection. It must not expose or retain raw prior answers, answer IDs, exact grades or scores, feedback text, or submission payloads. A projected own_response.record_present means only that a response record existed; it does not prove the student can view its contents. Keep own-answer visibility unknown unless separate safe evidence establishes it. If no safe projector exists, do not inspect the raw record: leave the capability unknown.

Label feedback evidence observed, declared, inferred, or unknown. Do not turn declared visibility into observed visibility. Two or more attempts plus useful feedback before retry plus keep_highest strongly promotes an otherwise suitable, mostly objective quiz. Merely allowing another attempt without timely useful feedback does not receive the same promotion. Never launch an attempt to discover any of these facts.