Skip to main content

MindRoom, Hermes and OpenClaw mesh

entry

Now​

Observed 2026-09-12, three repairs landed in the MindRoom mesh. Colima VM resources were raised from 2 CPU / 2 GiB to 4 CPU / 6 GiB (~/.colima/default/colima.yaml) while the 100 GB disk was kept. Before the change the guest reported 1,965 MB RAM with 13 MB available, zero swap, load average 121.97, memory PSI some avg60=96.53% / full avg60=51.90%, Docker inspection stalls and Matrix /versions plus authenticated timeout=0 /sync both timing out after 20s. After restart the guest reports 5,923 MB with roughly 3.9 to 5.1 GB available and memory PSI zero; Matrix HTTP recovered to 0.8 to 2.0s, sync timestamps advanced past the 120s watchdog interval, and both edge workers stayed healthy with zero uncertain jobs (completed counts Hermes 13, OpenClaw 4). One event-loop stall diagnostic at 22:01:51 measured 5.661 seconds and did not become a Matrix watchdog stall. No application code changed, so validation was live resource, database, transport and sync behaviour rather than unit tests. Source: /Users/dwayne/.hermes/reports/transport-pressure-20260912/README.md.

[HISTORICAL, MindRoom was removed 2026-09-19; these paths no longer exist.] The Matrix transport observer was rebuilt as a bounded design. chat.mindroom.transport-observer (launchd, StartInterval 30, KeepAlive false, program /Users/dwayne/mindroom/.venv/bin/python /Users/dwayne/.mindroom/bin/observe-matrix-transport.py) replaces the previous long-running loop with one bounded sample per invocation, so an exited sample no longer silently terminates all future monitoring. The tunnel byte cursor is preserved across invocations including truncation handling, host HTTP exceptions carry category and status, and the collector checks observation coverage against window boundaries with a maximum 60-second inter-sample gap. Scope ends 14 September 06:15 Sydney; later triggers exit without probing. Three regression tests passed (subprocess timeout, missing executable, sequential samples with host HTTP 503 and persistent tunnel cursor). The acceptance automation was reactivated for two new eligible windows, 13 to 14 September, then pause. Historical stall causation remains unresolved: the prior observer's exit code 120 alone does not identify why it exited, and a quiet window will not by itself prove a fix. Source: /Users/dwayne/.hermes/reports/mesh-repair-20260912/README.md, live plist and launchctl list read 2026-09-13.

Four ownership-warning source events have settled journal rows for all admitted principals, with router and final responding agent records completed; all eight referenced Matrix response events were independently read. Those receipts establish Matrix responses, not Telegram display or delivery. No replay was necessary or performed. Colima host-agent logs establish VM startup at 06:16:34 to 06:17 Sydney but do not establish why the VM had stopped. Source: same mesh-repair report (event-states.json, turn-outcomes.json, response-verification.json).

MindRoom context truncation was repaired separately. The reported Telegram message reached Matrix intact (2,321 body characters in source event $11eEuxvYQID63Mwv7FIktNp53OVQl8k8dr7LfhCFWhQ); AI-DLC's configured provider-context rendering capped each historical message to 200 characters, which explains the team's request to resend the supposedly truncated content. The 200-character per-message cap was removed while the existing 30-message and total token budgeting stayed. A shared tool-entrypoint guard now returns an explicit incomplete-result notice plus at most 64,000 preview characters for text strings above 1,000,000 characters, so the observed 15,905,473-character tool output against a 10,485,760-character limit cannot repeat on hooked entrypoints. 92 team-response tests and 44 tool-hook/result-size tests passed; Ruff, ty and git diff --check passed; MindRoom was restarted gracefully to load the code. Source: /Users/dwayne/.hermes/reports/aidlc-truncation-20260912/README.md.

Observed 2026-09-08: OpenClaw directly invoked the configured local executor /Users/dwayne/mindroom/ops/edge-workers/hermes_job.py for a bounded read-only Archify review. It returned ok: true, runtime hermes, an explicit acceptance/completion receipt and evidence-based findings; OpenClaw main retained ownership. This proves local executor-to-Hermes CLI execution, not MindRoom signed queue dispatch/enrollment/lease transport. The older documented ~/.openclaw/workspace/scripts/ask_hermes.sh and ~/.local/bin/fleet helpers are absent on this host. Source: OpenClaw task cd6ee682-0d9d-4501-b0e1-9e2f5dfa4390, exec process oceanic-slug, and /Users/dwayne/.openclaw/workspace/artifacts/archify-integration/evidence.md. <!-- project: path:/Users/dwayne/.openclaw/workspace -->

Observed 2026-09-08. Local host macbook-pro-3 (Zappo, 100.80.168.13) runs MindRoom on :8765, Hermes API on :8642 and OpenClaw on :18789. All return HTTP 200 from their respective ready/health endpoints. Pro (100.87.212.111) also runs an OpenClaw gateway with HTTP 200 health, verified through Tailscale SSH. Source: /Users/dwayne/.hermes/artifacts/agent-mesh/evidence.md and live probes in the originating Telegram session.

MindRoom's edge fleet has local Hermes and OpenClaw workers under LaunchAgents chat.mindroom.edge-hermes and chat.mindroom.edge-openclaw. The authenticated nodes API reported fresh heartbeats for hermes-local and openclaw-local. Each advertises chat and exec capabilities. Worker source implements signed enrollment, heartbeat, job leases, local CLI execution and result return. No end-to-end job was submitted in this audit. Sources: ~/mindroom/ops/edge-workers/{edge_worker,hermes_job,openclaw_job}.py, ~/mindroom/src/mindroom/api/edge_fleet.py, launchd and GET /api/edge-fleet-admin/nodes, observed 2026-09-08.

The Hermes-to-local-OpenClaw helper ~/.hermes/bin/openclaw-bridge calls the agent CLI with a topic session. This is a configured route, not verified successful turn execution. Source: that script, observed 2026-09-08.

MindRoom uses Matrix with Cinny and the mautrix-Telegram bridge. Synapse, PostgreSQL and Redis containers were healthy; Cinny and the bridge were running under Colima. Container health is not proof of bridge message delivery. Source: Docker process inventory and Matrix versions endpoint, observed 2026-09-08.

The knowledge layer is a local Git-backed fleet reference. Each runtime keeps separate state. This audit does not establish automatic memory synchronization across runtimes or hosts. Sources: knowledge-layer INDEX.md and partner policy, reviewed 2026-09-08.

Architecture artifact: ~/.hermes/artifacts/agent-mesh/mesh-v2.html; editable JSON and evidence.md are beside it. Eight nodes and eight relationships pass nine showcase checks with zero errors or warnings. Browser checks pass at four desktop sizes. Visual inspection covered dark 1440x900 and light 2048x1320. Source: Archify delivery output and mesh-v2.visual-check.json, observed 2026-09-08.

Log​

  • 2026-09-13 (weekly synthesis): Absorbed three 2026-09-12 repair reports that had landed as artifacts but not in this entity. (1) Resource pressure: Colima raised 2 CPU / 2 GiB to 4 CPU / 6 GiB, guest PSI zero, Matrix sync stable past the 120s watchdog interval. (2) Transport observer replaced with a bounded 30-second launchd sample; scope ends 14 September 06:15 Sydney; historical stall causation still unresolved. (3) AI-DLC 200-character context cap removed and an oversized tool-result guard added. Also verified live 2026-09-13: chat.mindroom.transport-observer loaded with exit 0, chat.mindroom.edge-hermes PID 91856, chat.mindroom.edge-openclaw PID 77460. Sources: /Users/dwayne/.hermes/reports/{transport-pressure-20260912,mesh-repair-20260912,aidlc-truncation-20260912}/README.md, live plist and launchctl list., hermes.

  • 2026-09-08: Added a successful bounded local Hermes executor review receipt, without claiming MindRoom queue verification. No configurations/security settings or services changed. Source: /Users/dwayne/.openclaw/workspace/artifacts/archify-integration/evidence.md. <!-- project: path:/Users/dwayne/.openclaw/workspace -->

  • 2026-09-08: Created from read-only source/configuration inspection and live local/remote probes. No services, credentials, schedules, or agent jobs changed.