Where the CARE build left off, what changed in three and a half weeks, and what's worth stealing from the last 30 days. Two Opus research agents, ~250 items, 9 sources.
ultracode in any prompt~/.claude/.env.Checkboxes save on this device — use it as a working list.
2 July, the Fable session. Your "Marketing strategy with Fable" brain dump became the CARE blueprint: a portable, always-on work system on a new homelab container — verified two-tier memory, Todoist, Syncthing, four scheduled jobs, Hermes as the phone front door. The session died mid-build when Fable usage ran out. The blueprint survives: CARE — system blueprint ›
jobs/ holds only a README.The blueprint parked herdr as "young, replaces tmux, no notifications." That call is now stale. Since the May safety scan it went from ~1.5k to 21,209 stars and v0.5 → v0.7.5, and the parts that were missing are exactly what shipped: a socket API with event subscriptions, a CLI for driving agents (herdr agent start / prompt / send-keys / wait), and an integration that reports real blocked/working/done state (herdr integration install claude).
The @tonbistudio post you sent is the proof of the pattern: Hermes supervising other agents through herdr's socket API — which is, almost word for word, the "hermes should be able to check in on active claude code sessions" line from your July brain dump. Their status skill is open-sourced as flock-check › — read-only, well-designed, and worth porting rather than writing our own cc-status script as the blueprint planned.
ctrl+b q; agents keep running.herdr --remote thin-client mode bridges clipboard image paste from the Mac — the exact Cmd+V-over-SSH gap cmux couldn't confirm.tailscale serve (you already run the tailnet). Floats the blocked agent to the top, pushes a notification when one needs input. github ›codex-raw and goes unrecognised (issue #1354). Fix: install the integrations; diagnose with herdr agent explain.ctrl+b — collides with tmux habits, and keybindings are a common complaint. Still pre-1.0.cc-status SSH-scraping script becomes a flock-check-style socket-API skill for both Hermes and Claude Code.herdr --remote when you need clipboard images). The planned careshot helper is probably unnecessary now.claude -p as designed; workflows are for interactive sessions, and the bundled /deep-research (adversarial claim-checking) is a working template for how the verified-facts jobs should verify.Type ultracode in a prompt (or /effort ultracode) and Claude writes a JS orchestration script executed in the background — session stays responsive, intermediate results live in variables not your context window. Caps: 16 concurrent agents, 1,000 per run. /workflows is a live progress TUI; press s to save a good run as a reusable command.
The bundled /deep-research workflow has agents adversarially cross-check each other's claims and drops what doesn't survive — a working reference implementation of your verified-facts memory idea.
Fact, first-party: code.claude.com/docs/workflows · needs v2.1.154+, all paid plans
From the 51k-star awesome-claude-code list: Code on Incus gives each agent an isolated machine (root, Docker, systemd) with automatic threat detection; compass adds cost-tiered subagents + guardrail hooks with Claude Code/Codex parity. Inference: on your Proxmox box, LXC-per-agent is the natural translation — isolation without giving up skip-permissions speed.
Trigger.dev ditched git worktrees for parallel agents: port conflicts, DB migration clashes, per-worktree service duplication, ~10GB disk across two worktrees on a 2GB codebase. They moved to GitButler virtual branches with file-level commit routing via a skill.
For your repos (vault, marketing ops — no dev servers, no migrations) worktrees are fine and this pain doesn't apply. It matters only if the codebase-access ask to Nanz turns into running the actual app.
124 memory-category MCP servers catalogued; nothing has won. The closest match to your design — Mwe-MCP, human-readable markdown wiki that self-organises overnight — has 5 stars. Read it for design ideas, don't adopt. Your approval-gated two-tier design remains ahead of the market; Supermemory's benchmarks are self-reported, Cognee is the only one with peer-reviewed results.
Mwe-MCP on GitHub · mcpservers.org
VaultSearch (BM25 + vectors + fuzzy titles, fully on-device) and Smart Connections now cover locally what qmd does. Mostly duplication for you — the delta is driving retrieval from inside the vault UI on your phone rather than only from an agent session. Also spotted: an Obsidian plugin that pushes note checklists to herdr agents as todos and ticks them off automatically — checklist-driven agent work straight from a note.
Avelina — "self-hosted personal AI assistant that lives in Telegram, persistent memory, voice, Claude-powered" — is the same shape as Hermes on CT103. Roundups now name Hermes Agent alongside OpenClaw, LibreChat, AnythingLLM as the standard self-hosted set. You're not on an island; the pattern won. The demo framing that resonates: texting your agent from a coffee shop instead of running back to the laptop.
coreyhaines31/marketingskills is the most-cited pack: CRO, copywriting, SEO, email/lifecycle, pricing, analytics — mapped across acquisition → activation → retention. Your bespoke CIO/PostHog/Strapi skills are better fitted than anything generic, but the funnel-stage map is a good audit lens against the 26 skills already in care-system: where are the gaps by stage?
Competitive signal from the same sweep: ActiveCampaign now ships 25+ agents (campaigns, automations, insights) — CIO's peers are putting agent surfaces into lifecycle tooling.
Portable agent config across Claude Code / Codex / Cursor / OpenCode is becoming its own category (ECC v2; a hackathon-winning repo with 28 subagents/119 skills doing 1M+ TikTok views). care-system's pinned-plugins + AGENTS.md-symlink approach is already the right idea — these are worth a browse for conventions, not wholesale adoption. Also flagged: obra/superpowers reported at 261k stars by the GitHub API — anomalously high, sanity-check before quoting.
~/Documents/Last30Days/: herdr-terminal-agent-multiplexer-raw-v3.md, claude-code-workflows-and-agent-setups-raw-v3.md, self-hosted-always-on-personal-ai-agent-setups-raw-v3.md.