← Reports
Release review · website docs · Luci architecture impact

Hermes v0.15 is less “chat app”, more operating system for agents.

The practical shift: durable work queues, safer restarts, faster search, richer skills/MCP, stronger gateway/Docker foundations. For us, it reinforces the split we already use: Telegram as control surface, Mission Control as source of truth, and Hermes cron as outside watchdog.

v0.15.0 Velocity Releasev0.15.1 Patch Release1,391 commits since v0.14Verified on Luci: v0.15.1
1,302
v0.15.0 commits
747
merged PRs in v0.15.0
28
v0.15.1 hotfix commits
20k
skills catalog scale after fix

Executive readout

The headline is not one single feature. It is a platform maturation wave: kanban becomes a real multi-agent substrate; cron/skills/MCP/dashboard get more operational; session search becomes cheap and instant; security/gateway/Docker become safer for unattended agents.

For Luci + MC, this means we should use Hermes primitives around Mission Control, not replace MC. MC remains the business workflow board. Hermes cron/watchdog and Hermes kanban become supporting infrastructure for agent operations, experiments, and specialist side lanes.

Status on Luci

Installed Hermes Agent v0.15.1 is live and current with origin/main.

Healthy Gateway, Telegram, cron, MC watchdog, and board shepherd are active.

Note Optional provider/tool warnings are expected and not blockers.

Feature map

Kanban becomes a multi-agent platform

Use

Auto-decomposition, swarm topology, parallel workers, gated verifier/synthesizer, per-task model overrides, task scheduling, worker visibility endpoints.

Our impact: Keep MC for business tickets; use Hermes kanban for specialist side work, experiments, and profile-native collaboration where MC overhead is too high.

Session search rewritten

Use

Moved away from aux-LLM dependency to fast FTS discovery/scroll/browse modes.

Our impact: Better recall before asking Elmar to repeat context. Use it for “what did we decide?” and operations archaeology.

Cron is more operational

Already using

Skill-backed jobs, script-only jobs, workdir/profile knobs, delivery routing, cron docs/troubleshooting.

Our impact: MC control-plane watchdog and board shepherd are the right pattern: quiet, verified, recurring, durable.

Skills ecosystem expands

Use more

Skill bundles, hub health, freshness badges, huge catalog indexing fix, better skill docs and pages.

Our impact: Bundle recurring Luci workflows: MC ops, research→artifact, email triage, audio briefings.

MCP catalog + mTLS

Selective

Interactive MCP picker, approved catalog, credential prompts, HTTP/SSE/TLS client cert support.

Our impact: Good for safely adding external tools. Do not add random MCPs to the core Luci profile without a reason.

Gateway/session resilience

Benefit

Topic binding fixes, media delivery fixes, restart/drain work, gateway log rotation recovery.

Our impact: Fewer Telegram/control-topic weirdness failures; still keep CCGram one-poller rule.

Docker/s6 supervision hardening

If containerized

Supervised gateway mode, persistent containers, orphan reaper, logs, UID/GID fixes, safer Docker update guidance.

Our impact: Not central on this bare-metal Luci setup, but useful if we spin isolated runtimes or packaged workers.

Security hardening

Benefit

Promptware scanning, memory/tool result delimiters, file-safety guards, Bitwarden Secrets Manager bootstrap, stricter API/webhook requirements.

Our impact: Move toward Bitwarden for stable secrets; keep auth/token paths protected from casual tool reads.

Dashboard + TUI maturation

Explore

Session control API, SSE chat, skill/toolset API endpoints, multi-session TUI orchestrator, dashboard polish.

Our impact: Useful for non-MC Hermes operations dashboarding; MC remains primary for business workflow visibility.

Context engine / compression plugin surface

Watch

Pluggable context engine docs, safer compression fallbacks, memory providers receive completed-turn context.

Our impact: Better long-running sessions; be cautious with cloud memory providers because completed-turn context can include raw content.

Dashboard 401 reload-loop fixed

v0.15.0 could reload forever in loopback mode because `/api/auth/me` 401 was expected but treated as stale auth. v0.15.1 fixes it.

Docker dashboard insecure mode explicit

Binding to non-loopback no longer implies insecure mode. Use `HERMES_DASHBOARD_INSECURE=1` deliberately.

MCP Docker command resolution

Bare `npx`, `npm`, `node` resolve correctly in Docker.

Kanban worker termination + image context

SIGTERM actually kills workers; images referenced in task bodies are attached to vision-capable workers.

Media/redaction fixes

`.md` file delivery restored; web URLs no longer get over-redacted in useful ways.

Model picker + `/yolo` fixes

`/model` and `hermes model` list alignment; `/yolo` now changes session bypass, not only env.

How this affects our architecture

Telegram topic

Control/status/escalation. Keep it lightweight; do not turn it into the workshop.

Luci / Orchestrator

Classifies intent: direct answer, delegate_task, MC ticket, Hermes cron, or background runtime.

Mission Control + Hermes

MC owns business workflow truth. Hermes provides outside watchdog, cron, skills, MCP, and optional kanban side lanes.

Keep Mission Control as the spine

Hermes kanban is now much stronger, but MC already encodes Elmar/Luci workflow, ticket URLs, runtime rows, board CTA, and Luci-specific review gates. Do not migrate the operational board just because kanban improved.

Use MC for
Elmar-visible tickets, Life/PA items, SafairBru/LegalMind work, Luci runtime accountability, review gates.
Use Hermes kanban for
Multi-profile experiments, plugin/tool spikes, auxiliary agent swarms, non-business durable subtasks.

Use Hermes as the outside observer

The new cron/gateway stability makes our watchdog pattern better: Hermes looks at MC from the outside and alerts only when the controller is unhealthy or a human decision is required.

Already live
MC control-plane watchdog every 15m, quiet unless alert condition.
Already live
MC board shepherd every 5m, local delivery, moves/gates routine board work.

Recommended adoption playbook

Now: codify the control-plane split

Telegram = control/status; MC = source of truth; Hermes cron = outside observer. Keep this as the default mental model.

Now: create skill bundles for Luci work

Bundles worth adding: mc-ops, research-html-report, email-triage, audio-briefing. This reduces repeated skill loading and steering.

Next: secrets cleanup

Evaluate Bitwarden Secrets Manager for stable provider credentials and workflow tokens. Keep Google/NotebookLM auth wall separate because browser sessions remain human-owned.

Next: MCP catalog as controlled intake

Before installing random external tools, check whether the official MCP catalog has a vetted server and install with explicit filters.

Later: use Hermes kanban for side swarms

For work not needing MC visibility, use kanban lanes. For Elmar-visible deliverables, create MC tickets.

What changes in our habits?

1
Search memory before asking again
Session search is faster and cheaper; use it proactively.
2
Prefer quiet script-only cron for watchdogs
No-agent watchdogs should output only alerts; empty output means healthy silence.
3
Keep provider auth explicit
xAI/GWS/NotebookLM auth failure detection improved, but human-owned OAuth still needs clear receipts.
4
Gate dashboard exposure
v0.15.1 makes insecure dashboard mode more explicit. Keep Tailscale exposure temporary and deliberate.

Risks / caveats

Kanban overlap risk

If we use Hermes kanban for everything, we split visibility from MC. Rule: MC for Elmar-visible business work; kanban for profile-native coordination and experiments.

Memory provider privacy

Completed-turn context can now be sent to memory providers. Keep built-in memory unless a provider has a clear privacy contract.

MCP tool sprawl

The catalog makes installation easier. Still require explicit purpose, filters, and rollback plan.

Dashboard auth assumptions

Docker/non-loopback exposure no longer implies insecure mode. Good for safety, but deployments must set auth/insecure flags deliberately.

Source trail

Compiled from official release notes and website docs after updating Luci to v0.15.1.