CLAUDE.md · git:20261027.32e5d1c · 2026-10-27 · sha256 d446cc0b28677fa8

CLAUDE.md git:20261027.32e5d1cA

Immutable. This exact content never changes and is served at /api/v1/blob/d446cc0b28677fa8.

# CLAUDE.md

Guidance for developing the Canvas MCP server. Agents *using* the server: see [AGENTS.md](./AGENTS.md).
Design: [internal/architecture.md](internal/architecture.md). Long-form rules and the reasons behind them: [internal/dev-reference.md](internal/dev-reference.md). Completed work: [internal/project-history.md](internal/project-history.md), `CHANGELOG.md`.
Codex reads `AGENTS.md`, not this file; its "Developing this server" section points here, so keep development rules in this file only. Paths, flags, symbols and issue states below were checked on 2026-09-30.

## Commands
- Install `uv pip install -e .`; run `canvas-mcp-server` (`--test`, `--config`); `.env` holds `CANVAS_API_TOKEN` and `CANVAS_API_URL`.
- Before committing: `uv run python -m pytest tests/ -v -rf`. TypeScript changes: `npm test` and `npm run build`.

## Git Workflow - ASK FIRST
- Before a feature or significant change, ask: feature branch or `main`? Defaults: new tool `feature/`, bug `fix/`, docs or typo may go straight to `main`. Prefixes: `feature/`, `fix/`, `docs/`, `refactor/`.
- The primary checkout stays on `main`, clean and read-only. One PR = one sibling worktree `canvas-mcp-<slug>` from `origin/main`; never reuse it for another issue; remove it and the branch (local and remote) after merge.
- After a sibling PR merges, rebase surviving worktree branches and rerun tests. Before tagging, run the suite in the primary checkout on the merged commit (its `.env` sets a token that worktrees and CI lack).
- `main` requires a PR plus `test-enhancements` and `lint`; the admin can bypass the review requirement.
- Run `./scripts/install-hooks.sh` once per clone. `fixes|closes|resolves #N` mid-sentence in a commit or PR body closes the issue on merge: rephrase (`closed [issue 172]`) or set `ALLOW_CLOSING_KEYWORD=1`. A trailer that opens a line (`Closes #173`) is allowed.
- Release steps and publish-race fixes: [internal/release-checklist.md](internal/release-checklist.md). Breaking changes need a minor bump.

## Coding Standards
- Type hints on every function; PEP 604 unions (`X | Y`), enforced by ruff UP.
- Tools use `@mcp.tool()` with `@validate_params`, and every tool carries annotations (`tests/test_tool_metadata.py` fails a bare decorator).
- All Canvas calls are async and go through `make_canvas_request()`; paginate every list endpoint.
- Course identifiers are `str | int` resolved with `get_course_id()`; dates go out through `format_date()`.
  Code that needs a numeric course ID (paths, context codes, comparisons) uses `resolve_numeric_course_id()`, which returns `(id, error)` and never yields an unvalidated string. `get_course_id()` matches the same way but keeps its pass-through fallback on a miss.
- Canvas POST/PUT needs `use_form_data=True`, `/conversations` included.
- Errors: dict tools return an `"error"` key; string tools return `"Error ..."`. `modules.py` and `accessibility.py` return JSON-stringified errors; keep the local convention.
- Privacy: student IDs preserved, names anonymized at the client layer (`_should_anonymize_endpoint()`); a new endpoint must be checked against the tier rules.
- Show course codes, not IDs, in user-facing output; handle published and unpublished states.

## Testing and behavioral evidence
- New tools need meaningful automated coverage: success, failure, boundary and safety behavior.
- A reproducible bug gets a failing regression first; watch it fail for the intended reason.
- Assert outgoing request contracts and forbidden side effects, from an independent requirement, not from the implementation's own output.
- Use a real client with controlled HTTP transport when the risk is serialization, retries, pagination or error classification.
- Never send live Canvas writes to verify a patch. Never weaken an assertion to get a green run.
- Lean/TLA+ (`verify/`) only for consequential state, ownership or termination invariants; pair with implementation regressions.
- Report what was verified, what failed and what was skipped.

## Documentation Maintenance
- Source of truth: agents `AGENTS.md`; humans `tools/README.md`; machine `tools/TOOL_MANIFEST.json`; `README.md` is the entry point.
- New tool: update `tools/README.md`, then `AGENTS.md`, then `TOOL_MANIFEST.json`. Touch `README.md` only for a major feature. No tool usage docs in this file.
- `docs/` is the public Cloudflare Pages root and deploys manually with `scripts/deploy_site.sh` (a bare `wrangler pages deploy docs/` ignores `.assetsignore` and publishes `triage/`, `audits/`, `superpowers/`); nothing internal goes there.
- The site's version, tool and skill counts and GitHub stats come from `scripts/sync_site_facts.py` (CI runs `--check`): change the source, then run it; never hand-edit those numbers.
- `internal/` is deny-by-default in `.gitignore`; un-ignore only a file meant to be public. `internal/session-history.md` stays untracked.
- Triage briefs are tracked: never record a collaborator's affiliation, evaluation status, timeline or competing products. Name the person and the technical issue only.

## Hosted Deployment
- A private, Entra-gated Azure instance serves Gies course staff. Its URL, app IDs, deploy details and key holders stay out of this repo; they live in the gitignored `internal/ops-hosted.local.md`.
- HTTP mode: each caller sends `X-Canvas-Token`; `CANVAS_API_TOKEN` must never be set (startup guard); read-only unless `ALLOWED_WRITE_TOOLS` is set.
- Merging to `main` does not deploy production. It deploys on a `v*` tag or `gh workflow run deploy-prod.yml --ref main`; `staging` deploys on push.

## Adoption numbers
- Never print a PyPI download count. Quote stars, forks and contributors, re-pulled from the GitHub API that day, with the date.
- Do not repeat "over 18,000 clones" or "UMich selected it as sole candidate" without a primary source.
- Do not imply a campus security review: the only Illinois artifact is Adam King's LRA, and the public hosted server is retired.

## Current Focus
- [ ] **#157** sandbox egress is mitigated, not closed (self-hosted only; `execute_typescript` is disabled on hosted). Needs an egress proxy or network namespace.
- [ ] **#236** OAuth2 developer-key flow: additive only, blocked on admin access to pilot a scoped key.
- [ ] **#172** Canvas Quizzes tools: blocked on a New-Quizzes-enabled sandbox (PR #191 was closed as unverifiable).
- [x] **#418 to #421** shipped (#438, #437, #436, #435; #421 part 2 via #442 and the anonymous-update refusal via #446; issue closed by 2026-10-07).
- [x] **v1.16.0** released 2026-10-24 (PR #530, tag on `b3f06c7`; Release + `.mcpb` with attestation, PyPI wheel + sdist with provenance, Registry latest, prod `azure-gies-b3f06c7` per the Kudu docker log, site from the tag plus current stats): security release; GHSA-hmr8, GHSA-7pp5, GHSA-7mj9 and GHSA-qf6c published 2026-10-25 04:11 UTC, CVEs requested (IDs pending). Breaking: the local sandbox needs Node.js 22.18+.
- [ ] **fastmcp** 4.1.2 is on `main` via #531 (merged 2026-10-25 18:08 CDT; #532 closed), which clears GHSA-7xgj-8392-5q4r and GHSA-vqqp-x5qm-5qp4; the dependency scan passes again. The released 1.16.0 (PyPI, `.mcpb`, prod) still locks 4.0.3 until a 1.16.1 release, which should also carry #535, #541 and #542.
- [ ] **OSS Verify** theorem-labs/oss-verify#8: Theorem's P3 change done (`326cb40`, compares the kept text; `--twice` passed), reply posted 10-24. **Merged** by `sjawhar` 2026-10-25 00:12 CDT (`66e6f43`; bot approval 10-24 23:40 CDT); `projects/canvas-mcp/` is on Theorem's `main`. Not in the proofs index, so say "enrolled", never "verified". The bot's post-merge review (10-25 00:59 CDT) listed 8 fuzz_html follow-ups; all 8 are in theorem-labs/oss-verify#42 (opened 10-25, 11 records met locally, `--twice` identical on m2max). **Theorem paused proof work on canvas-mcp** 10-25 20:02 CDT (no Python proofs yet, no date): #42 stays open, unmerged, no changes asked; enrollment stays. Their statements found a crash on a decimal entity of 4,301+ digits (CPython int-digit limit), fixed in #542. The stronger P2 check found the `_RAW_END` Unicode-folding bug, fixed in #535. Theorem statements first; proofs wait on their verified Python runtime. Findings go privately to vishal@illinois.edu.
- [ ] **OSS Scanner** anthropics/oss-scanner#1115: enrolled 2026-10-24 (terms and CLA signed at Vishal's direction, all checks green); waiting on Anthropic's human review.
- [ ] Hosted ops: remove `llkarri2` from the allowlist; usage telemetry is not retained (see Session Log).
- [ ] Watch, no owner: Agent Plugins packaging (blocked on credential delivery; triggers in project-history).

## Roadmap
- [ ] Grok/xAI integration research and pilot.

## Backlog
- [ ] Module templates; bulk module creation from JSON/YAML; module duplication across courses.
- [ ] Page templates; bulk page creation from markdown; page versioning tools.
- [ ] 2026-09-06 security review follow-ups: check `docs/superpowers/plans/2026-09-06-security-review-followups.md` for items PR #360 did not cover.

## Session Log
> Full history: `internal/session-history.md`, local-only and untracked since 2026-08-20. Do not re-add it to git.

### 2026-10-25 → 10-27 — OSS Verify follow-ups; html fixes; site simplified with fact sync; #541 merged
- **OSS Verify:** the 8 post-merge follow-ups went in theorem-labs/oss-verify#42 (11 records, `--twice` identical). Theorem **paused proof work on canvas-mcp** 10-25 20:02 CDT: #42 stays open, unmerged, no changes asked; enrollment stays.
- **html.py:** #535 (`_RAW_END` needs `re.ASCII`: `</ſcript>` and `</scrİpt>` ended script hiding) and #542 (a decimal entity of 4,301+ digits raised `ValueError` from `html.unescape`; `unescape_html` now guards all three call sites; 200,000-case fuzz against unlimited `html.unescape`, 0 mismatches). Both came from the independent fuzz checks.
- **Site:** #538 and #539: effects removed, impact panel static and moved before "Built for Your Role", Monthly Installs card removed, @vishalsachdev on X. `scripts/sync_site_facts.py` (CI `--check`) owns version, tool, skill and GitHub numbers; `scripts/deploy_site.sh` honours `.assetsignore` (Pages does not: `triage/`, `audits/`, `superpowers/` were live until 10-25). Cloudflare AI-training block kept zone-wide (Vishal, option B). The 10-26 scheduled impact run deployed itself (307 stars).
- **#541** (lindsay-cheng; peer-review campaign used the raw course code for its assignment GET and message context) merged 10-26 09:19 CDT as `1de55e4` with the prepared numeric-course fix; red-first 6 failures; 3798 passed; approval on head came from the prepared-fix reviewer on Vishal's account. Earlier runner SOCKS failures were its `ALL_PROXY` env (httpx needs `socksio`).
- **Next:** 1.16.1 release (fastmcp 4.1.2, #535, #541, #542; control owns it). Open PRs #545 (lindsay-cheng), #533 (KKazuhaK), #510 (mobile OAuth re-review). CVE IDs for the four advisories; Anthropic's review of oss-scanner#1115. Carry: F6 token revoke, remove `llkarri2`, #489.