Canvas Course Sync

Recurring job: sweep every Fall 2026 course on Canvas and pull anything not already local: module items, assignment attachments, announcements, files. Canvas API tokens are DISABLED at this school, so everything goes through the user's logged-in preview pane.

School config (READ BEFORE EVERY COURSE SESSION)

The suite ships with <PLACEHOLDER> tokens; resolve them from ~/.hermes/school-config.md (copy school-config.example.md there and fill in YOUR values once). At the start of any course session, read that file; if it is missing, tell the user to create it and stop — never answer with literal placeholders.

Student ID, school domain, course IDs, instructor/org names, platform URLs, and CLASSES_ROOT all come from the config. Never hardcode real values into any SKILL.md — that is what the config file is for.

To resolve values quickly, run python scripts/read_school_config.py (it reads ~/.hermes/school-config.md and exits 1 with a hint if missing).

Course IDs

Your school's IDs go in ~/.hermes/school-config.md (each course URL shows /courses/<ID>/). Keep a row per course, e.g.:

course_id_course1: <ID>
course_id_course2: <ID>
course_id_course3: <ID>

• The agent sweeps Canvas courses with the IDs from config; add a note next to each for quirks (disabled assignments, external homeworks, lab cadence).

Fresh-machine config (read me before first sweep)

• Classes root: this suite assumes a Classes/ folder (syllabi, materials, Roadmap.md, Announcements/). On a fresh machine, set CLASSES_ROOT (default C:/Users/<USER_NAME>/Desktop/Classes on the origin machine) and use it everywhere instead of the literal path.
• The one-time login in the pane (Dashboard → courses) is required; Achieve (Macmillan) and Pearson (MasteringEngineering) cannot be pane-logged — the user self-checks those and the agent reminds them weekly.
• Dependencies: python3 with faster-whisper pymupdf pillow, plus yt-dlp (+ bundled ffmpeg). See the suite README for install per OS.

Kaltura recorded lectures (download + transcribe)

Lecture videos are Kaltura playManifest URLs; the ks token in them EXPIRES within hours, so every video needs a FRESH URL from the user (F12 → Network → filter m3u8 → play → copy the a.m3u8). Never store the token — redact it.

Packaged pipeline (download + faster-whisper transcription, one background call):

cd "<Course>/Materials/Lectures"  # or wherever the video belongs
bash "$SKILL_DIR/scripts/kaltura_pipeline.sh" "<fresh a.m3u8 URL>" "<Output Base Name>"

• Outputs <Name>.mp4 + <Name> - transcript.txt + <Name> - transcript.md (timestamped [MM:SS] lines).
• Name by entryId lookup in scripts/kaltura_entries.tsv (entryId → canonical name); append new rows as you discover them so future grabs auto-name. entryId is the entryId/<id> segment of the URL.
• Run downloads in background with notify (terminal background=true + notify=true): 40–90 min lectures exceed the 420s foreground cap. yt-dlp -c -N 8 resumes safely across the restart.
• Verify each mp4 with C:/Program Files/yt-dlp/ffprobe.exe after download; transcription takes ~10–25 min per hour of audio on CPU (int8). A DONE <n> segments, <x> min line means the transcript is written.
• After transcription, skim the transcript for Key Points and file them beside it; log flags to Roadmap.md.

Pane driving (the only read path)

1. tool_call desktop_preview open https://<SCHOOL_DOMAIN> (user stays logged in; reopen if the pane was closed).
2. tool_call drive_preview elements first every page change; act by ref, not guessed selectors. Refs die on navigation — fresh elements after every navigate.
3. Read page text with tool_call desktop_preview read (wait ~2s after navigating).
4. Per course navigate to: /courses/{id}/modules, /courses/{id}/assignments, /courses/{id}/announcements, /courses/{id}/files (files page is often hidden; modules usually carry everything).

Sweep checklist (per course, in this order)

1. Announcements → compare against Classes/Announcements/<Course>.md (one md per course, ## YYYY-MM-DD — Title newest-first, full body). Append anything missing. If disabled (Circuits, Statics — verified), note and skip, don't re-check every time.
2. Modules → enumerate items. Files: download anything missing from local Materials/; external tools/videos: record existence + link in the course README (don't block the sweep on video extraction — flag for later).
3. Assignments → note new due dates for Roadmap.md + course README deadline tables. Download attached assignment sheets to Materials/Homework/.
4. Files → download anything not already local.

Dedupe + storage rules

• Local compare = same title OR same content; only pull what we don't have.
• Keep the professor's ORIGINAL filenames (they encode lecture order, e.g. <COURSE_CODE>-4a_).
• Multiple formats of the same lecture: keep PDFs only (user's standing rule for lectures).
• Layout: <Course>/Materials/ → Syllabus.* (original name), Lectures/, Homework/ (HW{n}.pdf), Homework Solutions/; textbooks at Materials/ root as <Author> - <Title> (<ed> ed).pdf.
• Student work + AI guides: <Course>/Materials/Homework/HW{n}/ holds the user's own work (scanned JPGs/PDF) plus any AI-chat export (e.g. Circuits HW1 AI Guide.md — Gemini) as a reference; file them there so every attempt/verification is recoverable next to the problem sheet.
• Downloads land in the user's Downloads folder via the pane; move them into the right course folder as part of the sweep. Never auto-process ~/Downloads for categorization, but class-material moves out of Downloads are expected.

Other grading platforms (user self-checks — assistant REMINDS)

The user checks both sites in their own browser (the pane can't hold those logins — confirmed 9/12). The assistant's job is the weekly reminder (Sunday review + before due windows) and logging what the user reports/screenshots. Read-only; never submit or change anything.

• Achieve (Macmillan) — homework platform. Course URL: https://<ACHIEVE_COURSE_URL>/mycourse (from config). The week view lists each HW with due dates + point values, review assignments, and adaptive quizzes.
  • Pane login does NOT work for Achieve (confirmed 9/12) — the user checks the week view in their own browser and reports/screenshots; log to Chemistry/README.md + Roadmap.md.
  • Capture: weekly HW names/due dates/point values + completion status → Chemistry/README.md + Roadmap.md (the Achieve assignment list mirrors the screenshot the user takes).
• Pearson MasteringEngineering — homework platform. Course menu URL: https://<PEARSON_COURSE_URL> (from config). Also reachable via Canvas → "MasteringEngineering" external-tool link (SSO via the pane's Canvas session; may need one user login the first time).
  • Pane login does NOT work for Pearson either — the user checks their homework list in their own browser and reports/screenshots; log due dates + scores to the Statics README + Roadmap. Statics Canvas assignments are disabled, so Pearson is the ONLY source of Statics due dates, and the user-check is mandatory, not optional.
• The user will keep taking screenshots of the Achieve week view — treat those as a first-class scan artifact (vision-read them like the 9/11 batch).

After the sweep

• Update Classes/Roadmap.md (deadlines, stress-balanced by-day plan — user wants load-tagged 🟢/🟠/🔴 scheduling, not raw deadline lists) and the course README deadline table.
• Report what was new vs already-local per course, and any videos/external tools that need follow-up.
• Ask the user to run this weekly or when Canvas shows unread/new-content counts.

Pitfalls

• Pane flakiness: use the retry ladder in references/pane-retry.md (elements refresh → wait → URL check → tab handling → hand to human for hover-⤓ downloads / third-party logins).
• Kaltura ks tokens die within hours — a URL that 404'd means it expired; ask the user for a fresh one. A URL from yesterday NEVER works.
• Automating the URL grab (browser_exec local=true) requires Chrome fully quit first — never kill the user's tabs; ask first.
• drive_preview intermittently fails with 'arguments is not valid JSON' or 'pointer input never reached the page' — navigation usually still happened; re-run elements() for fresh refs and re-read the URL before retrying.
• The user may close the preview pane mid-sweep; reopen it. If they close it in the middle of a critical flow, say so.
• AI policy varies by course (one course may allow AI with acknowledgement, another forbids it on graded work). Never ghostwrite graded written work.