diff --git a/.agents/validate.sh b/.agents/validate.sh new file mode 100755 index 0000000..dfd38c6 --- /dev/null +++ b/.agents/validate.sh @@ -0,0 +1,90 @@ +#!/usr/bin/env bash +# .agents/validate.sh — validation gate for the phased-execution pipeline. +# +# A unit is only moved to .agents/phases/complete/ if this script exits 0, +# and a phase is only COMMITTED if its final pass exits 0 — so this script +# is the last line of defense for "no phase commits unless 100% passing". +# It MUST run the project's FULL test suite (every test, not a subset) plus +# its linter, so a failure anywhere in the project blocks the commit. +# +# Adapt the checks below to this project's real test suite, linter, and +# coverage floor, then commit the result. +set -uo pipefail +rc=0 + +if [[ -f pyproject.toml || -f pytest.ini || -f setup.py ]]; then + python3 -m pytest -q || rc=1 + if command -v ruff >/dev/null 2>&1; then + ruff check . || rc=1 + fi +fi + +if [[ -f package.json ]]; then + npm test --silent || rc=1 +fi + +# --- project-specific gate: container security headers (Phase 03, audit M-4) -- +# Builds the real nginx image and asserts all four hardening headers on every +# response class (/, /index.html, a hashed asset). Needs a container runtime; +# if the host has none the check is skipped loudly instead of failing silently. +if [[ -x scripts/check-headers.sh ]]; then + if command -v podman >/dev/null 2>&1; then RT=podman + elif command -v docker >/dev/null 2>&1; then RT=docker + else RT="" + fi + if [[ -n "$RT" ]]; then + ./scripts/check-headers.sh || rc=1 + # check-headers.sh removes its container but not its scratch image tag; + # drop it here so repeated gates do not fill container storage. + leftovers="$("$RT" images --format '{{.Repository}}:{{.Tag}}' 2>/dev/null | grep 'homepage-headercheck-' || true)" + if [[ -n "$leftovers" ]]; then + printf '%s\n' "$leftovers" | xargs -r "$RT" rmi -f >/dev/null 2>&1 || true + fi + else + echo "WARN: no podman/docker available — SKIPPING scripts/check-headers.sh (security-header gate not exercised)" >&2 + fi +fi + +# --- repo hygiene (language-agnostic; runs in every project) ------------------ +# Scoped staging commits everything the phase touched, so a file that is not +# ignored gets committed — including raw log drops and key material. These +# gates make the phase fail instead. Fix by untracking, keeping the working +# copy: git rm --cached +if git rev-parse --is-inside-work-tree >/dev/null 2>&1; then + # 1. A tracked file that matches a .gitignore rule should never be tracked. + junk="$(git ls-files -i -c --exclude-standard 2>/dev/null || true)" + if [[ -n "$junk" ]]; then + echo "repo hygiene: TRACKED files match .gitignore rules (untrack with 'git rm --cached '):" >&2 + printf ' %s\n' "$junk" >&2 + rc=1 + fi + + # 2. Personal-data / key-material file SHAPES that must never be tracked, + # even in a project with no .gitignore rule for them. + DANGEROUS='(^|/)[^/]*\.log$|(^|/)[^/]*\.jsonl$|(^|/)\.env($|\.)|\.(pem|key|crt|cer|p12|pfx|keystore|jks)$|(^|/)(id_rsa|id_ecdsa|id_ed25519)(\.pub)?$|(^|/)(credentials|service-account|client_secret)([-a-z0-9]*)\.json$' + bad="$(git ls-files | grep -E "$DANGEROUS" || true)" + if [[ -n "$bad" ]]; then + echo "repo hygiene: TRACKED files look like log drops / key material (untrack with 'git rm --cached ' and add a .gitignore rule):" >&2 + printf ' %s\n' "$bad" >&2 + rc=1 + fi + + # 3. Project-specific content scan for personal data inside tracked files + # (client IPs, account ids, ...). Empty by default — fill it in per + # project, and allow-list documentation ranges (RFC 5737 / 4193 / 3849, + # 127.0.0.0/8, RFC1918) so fixtures do not trip it. Format: one ERE. + PII_PATTERNS="${PII_PATTERNS:-}" + if [[ -n "$PII_PATTERNS" ]]; then + hits="$(git ls-files -z | xargs -0 grep -IHnE "$PII_PATTERNS" 2>/dev/null || true)" + if [[ -n "$hits" ]]; then + echo "personal-data gate: TRACKED files match PII_PATTERNS — redact them:" >&2 + printf ' %s\n' "$hits" >&2 + rc=1 + fi + fi +fi + +if [[ $rc -ne 0 ]]; then + echo "validation FAILED (see output above)" +fi +exit "$rc" diff --git a/AGENTS.md b/AGENTS.md index 2271d4d..59f5c3c 100644 --- a/AGENTS.md +++ b/AGENTS.md @@ -17,8 +17,14 @@ Rules for any agent working in this repository. deployed artifact — never test `src/` directly). - `npm test` — the Playwright E2E suite (pinned `@playwright/test` 1.63.0; do not bump without re-checking the cached browser revision). - - For any nginx/Dockerfile change: `./scripts/check-headers.sh` - (Phase 03+). + - `.agents/validate.sh` — the pipeline gate: full E2E suite + the container + security-header check + repo-hygiene checks; must exit 0 before a phase is + committed. + - For any `nginx.conf` / `nginx-security-headers.inc` / Dockerfile change: + `./scripts/check-headers.sh` (added by Phase 03, run automatically by + `.agents/validate.sh` when a container runtime exists). It builds the real + image, serves it on `127.0.0.1:18099` and must print + `HEADER CHECK PASSED`; don't substitute a hand-rolled `curl` for it. ## Never - Modify `.agents/PLAN.md` except via an explicit user-approved change to the @@ -37,6 +43,36 @@ Rules for any agent working in this repository. - Static site: edits in `src/` → `./build.sh` → `dist/` (hash-cached names). New root-level assets are picked up automatically by `build.sh`; nested directories are **not** (keep JS files flat in `src/`). +- Terminal input is untrusted: never render typed input or `commandHistory` + with `innerHTML` — route it through `updateDisplay()`/`textContent` + (audit M-1, guarded by the markup-payload test in + `tests/e2e/02-terminal.spec.js`). The only `innerHTML` uses allowed in + `src/terminal.js` are static prompt literals, the `clear`/prank element + wipes, and the canned `neofetch` output (the `cmdText === 'neofetch'` branch — + every other command output goes through `textContent`). In + `terminal-vim.js`/`terminal-achievements.js` `innerHTML` is acceptable only + with static markup or `escapeHtml()`-escaped text. +- The fake terminal does zero network I/O: `curl`/`wget` render canned output + from `simulatedCurlOutput()` / `simulatedWgetOutput()` in + `src/terminal-commands.js` (pure string builders, no DOM, no network). Keep + any browser-side fetch call out of `src/` entirely (audit M-2, Phase 02). + Guarded by the curl/wget regression test in `tests/e2e/02-terminal.spec.js`, + which records every request whose URL falls outside `baseURL` from + `playwright.config.js` and fails the suite if the list is not empty. +- Security headers live in `nginx-security-headers.inc` (the Dockerfile copies + it to `/etc/nginx/snippets/security-headers.inc`) and **every** `location` in + `nginx.conf` `include`s it. nginx does not merge `add_header` across + configuration levels, so never rely on server-level inheritance and never add + inline security `add_header` lines to `nginx.conf` — put them in the snippet + (audit M-4, Phase 03). `server_tokens off;` stays in the `server` block + (audit L-2). `scripts/check-headers.sh` guards this rule: drop one `include` + from a location and the check fails for that path. +- Featured-projects content is test-locked: the `#projects` grid in + `src/index.html` (six cards in a fixed order — Brain of Reese first with its + Gitea source link, ServWaf last with no link because it is closed source) is + asserted by the `featured projects show the six updated cards in order` test + in `tests/e2e/01-page.spec.js`. Edit the cards in `src/index.html` and update + that test in the same change (Phase 05). - Terminal modules: `terminal-commands.js` (canned data), `terminal-achievements.js`, `terminal-vim.js`, `terminal.js` (core) — load order is fixed in `index.html` (all `defer`); keep dependencies pointing