[manager]

Production release authorized by Wolf on 2026-10-13. OffSec stays off main.

• AI Atlas: the header row holds ATLAS / STORY / LADDER as an ink-framed switch with icons and falling leaves behind the unselected tabs. STORY and LADDER render the article (with its nine-step decomposition board and a legend for the red and blue marks) and the vibecoding ladder (climb plus level modal). Words come from src/lib/aiAtlas/readings.json, written by scripts/ai-atlas/strip-readings.mjs; tab and legend labels are pushable page words.
• Widget: Copilot model moves from Claude Sonnet 5 to Sonnet 5.5 (already on staging since 2026-09-28).

Checks: git diff origin/main carries no OffSec lines; tsc --noEmit clean.

🤖 Generated with Claude Code

[github-actions]

Claude finished @manager's task in 23s —— View job

---

Review of PR #250: Atlas Story and Ladder tabs, widget on Sonnet 5.5

I reviewed the full diff in one pass. I didn't run tsc or the app, and I didn't read all 1390 added CSS lines or all 876 lines of readings.tsx. I sampled them for SSR safety, innerHTML use and font sizes.

Checked and fine

• git diff origin/main | grep -ic offsec returns 0, so OffSec is not in the diff.
• readings.tsx has no top-level window or document access. window.matchMedia is guarded by typeof window, and the other accesses sit inside functions and effects.
• I found no innerHTML or dangerouslySetInnerHTML in readings.tsx. Marked text is rendered from {t, m} runs, so there is no XSS path.
• The public-repo rules are respected. strip-readings.mjs reduces the source to visible words and ladder data, and logs/ is gitignored. I did not open readings.json to confirm it holds only rendered words.
• The new tab and legend labels were added to PAGE_TEXT_DEFAULTS, and the tab labels to SHORT_LABEL_KEYS.
• The CLAUDE_MODEL bump to claude-sonnet-5-5 matches the model ID.

Points to look at

1. Font sizes vs the passport. Most new rules use var(--fs-*), which I could not resolve to px. The literal sizes 16px, 14px and 12px are all on the allowed list. Check that the --fs-* tokens also resolve to allowed values, and that nothing falls below 8px.
2. Inline style in readings.tsx. Lines 723, 744 and 761 use style={{ ['--k' as any]: ... }}. This is the single-dynamic-property case that AGENTS.md allows, but the as any could be a typed CSSProperties cast.
3. Imperative DOM in readings.tsx. The board and the level modal build DOM with createElementNS, createElement('div') and createElement('button'). This is a port of vanilla JS. It works if everything is cleaned up in effects (the resize listener is). It is fragile with React 19 strict-mode double effects, so confirm there are no duplicate nodes on remount.
4. strip-readings.mjs (inline comment): vm.runInNewContext is not a sandbox, and the lazy regex in literal() can truncate nested literals.
5. Tests. Only Playwright E2E applies in this repo. I saw no new test for tab switching or for the #hash sync that ai-atlas.tsx now handles for the new tabs.

No blockers found. Items 1 and 3 are worth a manual check on staging before the production release.

[github-actions]

Minor: vm.runInNewContext is not a security boundary. It's fine for a local, trusted source folder, but add a one-line note in the header that the script must only be pointed at a trusted checkout. Also, literal() uses a lazy regex ending at the first ]\n or }\n, so a nested literal that closes on its own line will truncate silently. The throw on parse failure covers most cases, but a count check against the expected step count would be safer.

[manager]

Review read, no change needed before merge:

1. The --fs-* tokens resolve to 16, 20, 24 and 32px, all on the passport; nothing is below 12px.
2. --k/--i are single dynamic custom properties, as AGENTS.md allows.
3. The board effect clears the stage, the step bar and the tip box both before drawing and in its cleanup, so a strict-mode double run leaves one copy; the level modal is plain React.
4. literal() cannot truncate silently: a cut literal is unbalanced and vm throws, which the script journals as a failure. The source folder is a trusted local checkout.
5. No Playwright test added in this release.

🤖 Generated with Claude Code