CLAUDE.md — Project memory for Claude Mastery Hub

Checkpoint: OMEGA · last updated 2026-07-12 Read this first. It's the durable brief so any new session starts grounded.

Session protocol (read this first)

Every coding session must read docs/CURRENT_STATE.md first and update it last. It is the single source of truth for branch, HEAD, active PR, CI run, test count, branch protection, and the one current next action. If it disagrees with memory or an older document, it wins. Supporting documents: docs/PROJECT_ROADMAP.md (plan and invariants) · docs/DECISION_LOG.md (why, and what it rules out) · docs/CHANGELOG_EXECUTION.md (what shipped, with evidence) · docs/roadmap/index.html (offline visual dashboard).

main is protected (enforce_admins: true). Every change — including a one-line documentation fix — goes through a PR with a green Quality gate. Direct pushes are rejected server-side.

What this is

An interactive, single-file educational guide to mastering Claude (the app, Claude Code, the API & Agent SDK, MCP, skills, subagents, hooks, plugins, multi-agent workflows) plus an "AI Foundations 101" track.

• Live page: https://gesh75.github.io/claude-mastery-hub/ (GitHub Pages)
• Repo: gesh75/claude-mastery-hub (owner: @gesh75)
• Everything ships in index.html — inline CSS + JS, zero dependencies, fully offline. No build step. open index.html just works.
• Other tracked files: README.md, og.png (1200×630 @2× social card), this file.

Architecture (all inside index.html)

• Layout: CSS grid .layout{grid-template-columns:288px 1fr}; collapsible sidebar via html.nav-hidden; mobile breakpoint @media(max-width:920px).
• Sections: <section class="sec" id="..." data-id="...">. Six nav tracks: Foundations · Claude Code Core · Power Features · Mastery · Practice · AI Foundations. Reveal-on-scroll via IntersectionObserver (.sec→.sec.in).
• Data-driven learning engines (near end of file):
  • QUIZ — per-section quizzes keyed by section id.
  • EXAM — the 20-question final exam (#exam), ranked tiers.
  • LAB — the Practice Lab (#lab): 25 hands-on scenario challenges, two types (type:"cmd" typed answer with accept/startsWith matching, or type:"choice"); each has hint, sol/solText, and a why note. "Show solution" unlocks only after an attempt.
• localStorage keys: claude_hub_quiz_v1, claude_hub_exam_v1, claude_hub_mastery_v1, claude_hub_nav_v1, claude_hub_lab_v1. Spaced-repetition intervals: [1,3,7,21,45] days.
• Mastery ring counts .master-cb checkboxes (currently 43); adding a section with a master-cb auto-increments the /N total — update the hardcoded ringmeta fallback text to match.

Conventions & hard-won gotchas

• NEVER name a JS variable top. It collides with the read-only window.top global; the assignment silently no-ops and top.onclick=... installs a window-wide handler (this caused the "every click jumps to the top" bug, fixed in #6). Use topBtn, etc.
• Building JS via a Python script? Don't put emoji in a Python raw string. \U0001f4a1 passes through literally and JS mis-parses \U. Use literal emoji characters, or 💡 surrogate pairs. (Bit us in #7.)
• Prefer atomic Python edit-scripts that read the file fresh and assert s.count(old)==1 before replacing — the Edit tool intermittently trips on "file modified since read" from an external toucher.
• Verify every change in headless Chromium before shipping: executablePath:'/opt/pw-browsers/chromium-1194/chrome-linux/chrome', playwright-core lives in the scratchpad. Check: renders, no console errors, no leaked U0001… escape artifacts, scroll preserved on plain clicks.
• Facts must be verified before writing. For anything about current Claude Code / models, check the official docs/changelog (code.claude.com) — don't write from memory. Models are fast-moving.

Current model lineup (verified 2026-07-28 — re-verify before editing)

Two Opus 5 migration exceptions (verified 2026-07-28): the API Web Fetch tool is not available on Opus 5 (Web Search is a different tool and is fine), and Priority Tier does not support Opus 5 (nor Sonnet 5) — and new Priority Tier commitments are no longer sold, so only existing commitments are affected. Neither limits ordinary Opus 5 availability on the API, Bedrock, Google Cloud, or Foundry. Fast mode is Claude API (incl. Managed Agents) + Claude Code via usage credits only — never Bedrock/Google Cloud/Foundry/Claude Platform on AWS — and its "up to 2.5×" is output tokens per second, not TTFT and not an SLA. 300k output is a Message Batches-only cap; synchronous Opus 5 stays at 128k.

/model default is plan-dependent. Subagents run in the background by default and nest up to 5 levels. Agent teams are experimental (CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1).

Ship flow (what has worked every time)

1. Branch claude/product-gaps-improvements-h5gdn4 (restart from origin/main after each merge: git fetch origin main && git checkout -B <branch> origin/main).
2. Edit → verify in headless Chromium.
3. Commit → push (--force-with-lease if the branch still carries a just-merged pre-squash commit).
4. Open a draft PR, then squash-merge on the user's OK.
5. Confirm the deploy. Fastest check (verified 2026-07-29): gh api /repos/gesh75/claude-mastery-hub/pages/builds/latest returns status: built plus the deployed commit SHA directly — compare it to the squash commit. The actions_list/gh api .../actions/runs?event=dynamic route still works as a cross-check ("pages build and deployment", matching head_sha, conclusion=success); parse large run JSON with jq/Python, not raw output.
6. Direct github.io fetch works (corrected 2026-07-29 — the older "blocked by the proxy" note was wrong). curl the live URL and diff its sha256 against the repo's index.html; identical hashes prove the deploy served this commit. First fetch can take ~25s. Playwright can also drive the live URL directly via test.use({baseURL}), which is how production smoke tests are run.

History (merged PRs)

#1 Sonnet 5 + How-To/CLI Hacks + animation explainers · #2 mid-2026 currency · #3 learning engine (quizzes/exam/spaced-rep) · #4 full quiz coverage + search + anchors + print · #5 collapsible sidebar · #6 fix global click-to-top bug · #7 Practice Lab (12 challenges) · #8 currency pass (Chrome, agent teams, Sonnet-5 default) · #9 og.png featuring the Practice Lab · #10 this CLAUDE.md (Omega memory) · #11 Fable 5 added to the "Pick the right model" quick table · #12 CLAUDE.md checkpoint (PRs #10–#11) · #13 P0 responsive containment + mobile nav a11y · #15 Copilot custom instructions · #14 Opus 5 currency: current lineup, MODEL_FACTS registry, provenance-labeled benchmarks (ARC-AGI-3 as 30.16% exact / 30.2% rounded), migration exceptions (Web Fetch, Priority Tier), Fast-mode scoping, content-currency policy, and Playwright currency regressions.

Merged branches are normally retained here (fix/p0-responsive-navigation, claude/product-gaps-improvements-h5gdn4, content/opus-5-currency all still exist); only #15's branch was deleted. Don't delete one without being asked.

Known loose ends / ideas

• Periodic currency upkeep as Claude Code ships (MODEL_FACTS last verified 2026-07-28; review-due after 30 days — see GAP_ANALYSIS.md).
• More Practice Lab challenges; harder "apply/analyze" exam questions. Advertised section / diagram / Lab / track counts are derived at branch time and guarded by tests/e2e/content-counts.spec.js (closed in #22 / #39).