Canvas Submit

canvas-submit

Perform one explicitly authorized Canvas mutation after the scan, approval, and draft boundary. This skill is the interactive submit entry point; neither canvas-scan, canvas-execute, nor a course drafting skill may invoke it on the user's behalf.

1. Enforce the later-message boundary

Use the raw text of the current user message. Never reconstruct authority from an earlier message, quoted text, a plan decision, an overlay, a schedule, or the agent's own recommendation.

Pass the complete current message to src.mutation_approval.parse_mutation_command. Accept only one whole expression recognized by that parser:

• submit N or 提交第N项 for one ordinary assignment;
• take quiz N, 参加测验第N项, or 做测验第N项 for one first quiz attempt;
• retake quiz N, 重做测验第N项, or 重考测验第N项 for one explicitly authorized retake.

Outer whitespace may be retained, but no other prose is allowed. Reject submit all, approve all, please submit 1, submit 1 and 2, take quiz 1 now, hypotheticals, and any command copied from a prior turn. One message authorizes at most one indexed target. Never translate a vague request into an accepted command.

Plan approval is not submission authority. approve 1, all, do 1, and a plan.json decision may authorize drafting through canvas-execute, but they must never mint or substitute for a mutation receipt. This skill is not an automatic-submit or cron workflow.

2. Bind the current plan target

Use only the run directory associated with the current scan presented in this conversation. Require its non-expired plan.json and matching assignments.json; do not search older run directories for a convenient target. Validate them with src.run_state.validate_plan_assignments(..., run_dir=run_dir, require_current=True).

The requested index must resolve to one contiguous 1-based plan item whose user_decision is approve or swap:canvas-*, and that item must have a one-to-one snapshot match. If the plan is absent, stale, unapproved, or inconsistent, stop and require a fresh scan/approval. Never repair plan state inside this skill.

Derive the only allowed work directory from immutable IDs:

from src.run_state import stable_work_dir

work_dir = stable_work_dir(run_dir, course_id, assignment_id)

It must be named course-<course_id>__assignment-<assignment_id>. Never use a course name, assignment name, slug, or arbitrary path as identity.

3. Require a verified ordinary-assignment draft

For submit N, perform this local fail-closed preflight before issuing a receipt:

1. Require <work_dir>/result.json to exist.
2. Load it and call src.run_state.validate_result(result, root=Path.cwd(), work_dir=work_dir); do not trust a status string without validation.
3. Require the validated status to be exactly draft_ready.
4. Require the resolved draft_path to be substantive and contained inside the stable work directory. Never submit a path supplied only in chat.
5. Call src.run_state.validate_verification_log on the verification evidence selected by the result contract. It must contain at least one PASS line, no FAIL line, and no unresolved placeholder/skeleton sentinel.
6. Confirm the result, plan item, snapshot, course ID, assignment ID, and intended payload all refer to the same target.

A missing draft, directory escape, error/skipped result, failed verification, or ambiguous payload must stop before Canvas mutation. Do not regenerate, edit, or silently replace the draft in this skill; return to the responsible course skill instead.

Quiz commands do not use this ordinary-draft gate. Their read-only reading, notes, arbitration, pacing, and mutation gates remain owned by canvas-inside.

4. Issue one exact signed receipt

After the applicable local preflight, call src.mutation_approval.issue_interactive_authorization with:

• the exact current run_dir;
• the configured origin from src.canvas_client.BASE, never a URL invented from chat;
• user_text equal to the verbatim current user message;
• the default protected local signing key and current Codex thread/session.

The helper repeats current-plan validation, derives the exact target/actions, and writes these private artifacts below the stable work directory:

mutation_authority.json
mutation_authorization.json

Use synthetic_qa=True only inside an independently verified, isolated local QA Canvas. It is always false for a real Canvas origin. Do not print, copy, publish, broaden, refresh, or hand-edit a receipt. A receipt for one course, assignment/quiz, origin, session, or action cannot authorize another.

If authorization issuance fails, stop without mutation. Receipt issuance by itself is not receipt consumption.