name: cortex-help
description: One-screen menu of what cortex-x can do right now — list of invokable slash commands (cortex-init, cortex-help, cortex-doctor, audit, designer, start, test-audit) + operator-facing terminal CLIs (cortex-update, cortex-uninstall, cortex-hooks-register, cortex-claude-md-augment, cortex-doctor), a 1-line summary of what each does, and a count of underlying capabilities (action_kinds, profiles, standards). Auto-discovered after install.{sh,ps1} sync. Invoke as "/cortex-help", "what can cortex do?", or when the user asks "co umíš?" / "jaké jsou skills?" / "ukaž mi capabilities". NOTE: namespaced as cortex-help because /help is reserved as a Claude Code built-in slash command. Defers full machine-readable registry to cortex/capabilities.md.
disable-model-invocation: false
/cortex-help — what can cortex-x do right now
Voice charter: see standards/voice.md. The menu IS the voice surface — no greetings, no emoji, no emotion words. Counts not praise.
You are running the cortex-x help menu. The user wants a one-screen answer to "what can I type next?" — NOT a wall of documentation. Print the menu, suggest a sensible default action based on detected project state, stop.
Step 1 — language signal
Read the user's language from prior turns. If they wrote Czech, answer in Czech. Otherwise English. Don't ask explicitly.
Step 1.5 — surface install state (one line, always)
Before the menu, print ONE line showing where cortex-x is installed. Read ~/.claude/shared/cortex-source.yaml via Bash or Read tool to get the resolved paths. Format:
cortex-x: source <cortex_source> · data <cortex_data_home>
Fail-open: if the yaml is unreadable, fall through to "(install state unknown — try cortex-update --check)".
This is load-bearing because the operator's friend / new-PC scenario commonly opens with "kde je cortex?" — surface the answer at the top of the menu so they don't have to ask.
Step 2 — detect current state (one quick scan)
Before printing the menu, peek at $PWD:
.cortex-bootstrap-pending exists → bootstrap was started, point them at the next stepcortex/AUDIT.md exists → audit already ran; suggest /retrofit nextcortex/discovery.md or cortex/proposal.md exists → new-project in flight; suggest resumingpackage.json / pyproject.toml / Cargo.toml / go.mod present → existing project; default = /audit- Folder near-empty → new project; default =
/cortex-init
You do NOT block on this — if any check fails, fall through to "no detection signal."
Step 3 — print the menu
Print this menu, with the "default" tag on whichever option Step 2 indicated. Format as table-or-list, choose what renders cleanly in the user's chat surface.
Slash commands you can type right now
| Command | What it does | When to use |
|---|
/cortex-init | Interactive picker — New / Existing / Framework — primes the right flow | Just installed cortex-x; not sure what to do |
/start | New-project bootstrap (Discover → Research → Architect → Scaffold → Adapt) | Empty folder, want a full project scaffold |
/audit | Existing-project deep audit (12 dimensions, 4 parallel agents) | Established codebase, want to know what's there |
/test-audit | Senior-QA-consultant audit → P0/P1/P2 gap list with research memos | Repo needs a testing-strategy review |
/improve-codebase-architecture | Read-only audit for shallow/zero-test modules → deepening candidates (AI-codeability lens) | Agents write bad code in one area; before a big refactor |
/designer | Designer flow — intake + library palette + parallel worktree exploration | Front-end / landing-page / dashboard design session |
/cortex-doctor | Healthcheck — cortex install integrity + drift detection (Sprint 2.21) | Weekly, after migration, or new machine |
/retrofit † | Apply cortex-x patterns to an audited project | After /audit finishes |
/sync † | End-of-session knowledge capture — decisions + lessons → cortex library | After a sprint or notable work session |
/cortex-reflect † | Deep reflection — surfaces grounded cross-project insights | When something feels off, or after big refactor |
/cortex-help | This menu | Anytime you forget what's available |
† = paste the prompt manually from ~/.claude/shared/prompts/<name>.md (these are prompts, not slash-skills yet; SKILL.md wrappers planned for Sprint 2.18+).
Terminal CLIs — run in shell, outside Claude Code
These are operator-facing maintenance tools shipped under ~/.claude/shared/bin/ (add to PATH or invoke directly). Run them in a normal terminal, not inside the Claude Code chat surface.
| Command | What it does | When to use |
|---|
cortex-update | Fast-forward source clone + re-run installer. Flags: --check, --reinstall, --json | Bump framework to latest |
cortex-uninstall | Conservative removal — preserves $CORTEX_DATA_HOME by default. Flags: --dry-run, --purge, --backup, --keep-source, --yes | Removing cortex-x from machine |
cortex-doctor | Same checks as /cortex-doctor slash skill but headless. Flags: --json, --fix-suggestions. Exit 0/1 | CI / scripted health audit |
cortex-hooks-register | Opt-in idempotent merge of cortex hooks into ~/.claude/settings.json with timestamped backup. Flags: --apply, --remove, --status, --dry-run, --yes, --json | Manual hooks wiring if install consent declined |
cortex-claude-md-augment | Opt-in BEGIN/END-marker append of discipline block to ~/.claude/CLAUDE.md (R1+R2+task-list+voice+surgical). Flags: --apply, --remove, --status, --dry-run, --yes, --json | Manual global-CLAUDE wiring if install consent declined |
cortex-permissions-register | Opt-in idempotent merge of safety-floor permissions into ~/.claude/settings.json (deny floor + allow baseline). Replaces --dangerously-skip-permissions: same speed, deny > ask > allow precedence floor. Flags: --apply, --remove, --status, --dry-run, --yes, --json | Manual permissions wiring if install consent declined |
cortex-usage | Usage rollup over the tool-use journal — which standards / prompts / agents / skills actually get read in sessions (hot = earns its context cost) vs cold prune candidates. Flags: --since, --kind, --cold, --json | Regaining overview · feeding the usage-driven pruning audit |
Full CLI roster (15 tools incl. internal Steward + lessons + propose-skill + evolve-ab + wiki-consolidate + doc-audit): see bin/README.md.
Default behaviors (these run without being asked)
| Behavior | When | Override |
|---|
| Web research | Whenever a task depends on external state (framework versions, library APIs, CVEs, design trends, a11y standards) — Claude is expected to dispatch parallel research subagents first, cache under $CORTEX_DATA_HOME/research/, cite URLs | --no-research flag · CORTEX_OFFLINE=1 env |
| Session-start context injection | Every Claude Code session — sprint state, git state, capability tip (once / 18h), pending insights | uninstall hooks or CORTEX_HOOKS_DISABLED=1 |
| Block destructive | rm -rf /, force-push to main, DROP TABLE, --no-verify are refused | run outside Claude Code |
| Auto-orchestrate review | Multi-agent parallel review pipeline auto-dispatches on non-trivial diffs | CORTEX_AUTO_ORCHESTRATE_DISABLED=1 |
What's under the hood (read-only summary)
The above are user-facing skills. Underneath, cortex-x ships:
- Steward action_kinds — typed actions the autonomous nightly runtime can execute (
dep_update_patch, recommendation, pattern_transfer, senior_tester_review, etc.)
- Steward primitives — zero-deps CJS modules implementing safety, dispatch, memory layers
- Universal hooks — block-destructive, session-start, pre-compact, post-tool-use, etc.
- 26 standards across 5 tiers: Rule 0 Ship-Ready → Rule 1 SSOT/Modular/Scalable → Rule 1.5 Coding Behavior + Self-Correction + Auto-Optimization → Rule 2 Security/Testing/Observability/Correctness → Rule 3 process (Accessibility/Performance/Errors/Git/Docs/AI-Patterns/AI-SDKs/Web-Research/Voice/Skills). Voice charter (
standards/voice.md) governs every skill's tone — no greetings, no emoji, no emotion words, [cortex/recall] + [^cN] footnotes for memory recall.
- 11 profiles — project-type templates (nextjs-saas, waas-template, chatbot-platform, ai-agent, browser-agent, kiosek, qa-engineer, tauri-desktop, astro-static, cli-tool, minimal)
- Review-pipeline agents — blind-hunter, edge-case-hunter, acceptance-auditor, security-auditor, ssot-enforcer, correctness-auditor, cortex-thinker
- GitHub workflows —
steward-*.yml cron lanes per action_kind
Full machine-readable registry with exact counts + descriptions: cortex/capabilities.md (auto-generated; refresh via npm run capabilities).
Don't see what you want?
Step 4 — single-line nudge
Below the menu, print ONE line: the suggested next move based on Step 2 detection. Examples:
- "Detekoval jsem
package.json v této složce → doporučuji /audit."
- "Folder vypadá prázdně → doporučuji
/cortex-init nebo přímo /start."
- "Vidím
cortex/AUDIT.md z předchozího běhu → /retrofit by byl logický další krok."
- "Žádný signál — vyber si z menu výše."
Step 5 — stop
After the menu + nudge, end the turn. Do NOT auto-invoke another skill. The user picks what they want.
Anti-patterns
- Don't dump capability registry verbatim. That's
npm run capabilities territory; this skill is the human-facing menu.
- Don't explain how each command works in detail. One line per command, full docs are linked.
- Don't repeat the README. README is for first-time discovery;
/help is for "I forgot what's here."
- Don't ask follow-up questions. Print, nudge, stop.
Composes with
/cortex-init — natural follow-up if user is fresh-installedcortex/capabilities.md (Sprint 2.15 auto-registry) — the machine-readable detailed view this skill links to
