"Distribution expansion — root CLI .sh commands (install/update/changelog/help) + remote MCP connector transport for agent UIs"

§1 — Description

1. update.sh, changelog.sh, help.sh MUST exist at the payload root beside the already-shipping install.sh, directly runnable (bash update.sh), mirroring the plugin commands' semantics: update = read-only check by default with --apply to execute (thin wrapper over install.sh, which owns the logic); changelog = installed version + rules_sha + pointers, read from the manifest beside the script; help = the command surface.
2. build.sh MUST ship the trio in every payload (copy + chmod + npm files array) and declare the two new channels in manifest.yaml: root-cli, mcp-connector.
3. cyberos-mcp.mjs MUST gain a --http [port] mode: MCP streamable-HTTP style endpoint (POST = one JSON-RPC message or batch → application/json; notifications → 202 empty; non-POST → 405; GET /healthz probe), reusing the existing handle() dispatch verbatim so stdio and http can never drift. stdio stays the default; zero new dependencies.
4. A docs/deploy/mcp-connector.md runbook MUST capture the agent-UI hookup (Claude: Name + Remote MCP server URL + optional OAuth; Grok: Name + Server URL) and the production checklist: reverse-proxy TLS, supervisor, and the explicit warning that the transport ships unauthenticated - public exposure requires proxy-level auth since the tools execute repo workflows.
5. Grok's dialog placeholder suggests legacy /sse; whether it accepts streamable HTTP is confirmed at hookup time, not asserted - if legacy SSE is required, that transport is a recorded follow-up (§9).

Length note: sanctioned lean profile - every §5 check ran live in-session (scripts executed, endpoint curl-verified).

§2 — Why

update.sh wraps install.sh instead of duplicating logic (single source of truth for vendoring + --check). The http mode reuses handle() so a tool added once serves both transports. Auth is deliberately NOT hand-rolled into the node process: proxy-level auth at nginx is the checklist item - a bespoke token check in a zero-dep server is worse than the battle-tested proxy layer already fronting the VPS.

§4/§5 — Acceptance + verification (all run live 2026-07-13)

1. Payload contains executable trio; check-version-sync.sh still green. ✅ (build to /tmp, ls + run)
2. help.sh prints the surface; changelog.sh prints version + rules_sha from its own manifest; update.sh (no args, offline) emits install.sh --check's three-value report + verdict. ✅ (all executed)
3. --http: /healthz 200 JSON; POST tools/list returns the 4 workflow tools; notification → 202; GET → 405. ✅ (curl-verified)
4. stdio unchanged (default branch untouched semantics). ✅ (code path conditional on --http only)
5. Runbook exists with hookup + security checklist. ✅

§5b — Testing pass (2026-07-13, post gate-1 "approve all")

• Payload rebuilt to /tmp: help.sh 17-line surface exit 0; changelog.sh prints 1.0.0 + rules_sha + pointers; endpoint healthz 200 / tools-list 4 / notification 202 / GET 405. PASS
• DEFECT CAUGHT + FIXED IN-PHASE: update.sh --check passed the flag through twice (exec install.sh --check "$@" without shifting), so install.sh read the second --check as its TARGET ("cd: --: invalid option"; root detection broken, installed=none from an installed repo). The no-arg default path had masked it at implementation time. Fix: case/shift in update.sh. Re-run, all arg shapes green - live report: installed=1.0.0 payload=1.0.0 latest=1.0.0 (GitHub releases reachable) verdict=up_to_date.

§9 — Open questions

• Grok legacy-SSE transport if streamable HTTP is rejected at hookup (recorded, not assumed).
• OAuth client id/secret support in the connector dialog - relevant only if Stephen wants Anthropic-managed auth instead of proxy auth; follow-up.
• Public URL + reverse-proxy wiring on the VPS (operator step - checklist in the runbook).

§10 — Failure modes

End of TASK-IMP-076.