Skip to main content

2026-07-09 — Fleet upgrade: bridge, shared memory, fallbacks, security trim

entry

Full detail: ~/.hermes/SETUP-CHANGES-20260709.md (fleet-core git).

Context​

Hermes and OpenClaw coordinated only by passively reading each other's Telegram output (both sides [SILENT] to bot messages); everything rode one model with no fallbacks; command_allowlist blanket-allowed destructive ops; instruction files synced by hand with .bak sprawl.

Decisions​

  • Direct bridge both ways (openclaw-bridge / ask_hermes.sh); group chat = human-visible only.
  • Knowledge graph declared the fleet event spine; kg wrapper for Hermes-side writes.
  • OpenClaw fallbacks: glm-5.2:cloud, then claude-fable-5 (Anthropic deliberately NOT re-added to Hermes — Dwayne removed it 2026-07-05, .env.bak-clear-anthropic-token).
  • Hermes aux quality: vision→gemini-3-flash-preview, compression/curator→glm-5.2 (memory quality compounds).
  • Security: 11 destructive allowlist categories removed; tirith fail-closed. approvals.mode left allow (Dwayne's call to change to allowlist).
  • Instruction layer under git (fleet-core, ~/.hermes, default-deny gitignore, 16 files).
  • Live gateway-managed files (jobs.json, exec-approvals.json) are staged-only, never hand-edited.
  • Knowledge layer (this repo) added same day as the synthesized-truth store.

Rollback​

*.bak-fleet-upgrade-20260709 for both configs; fleet-core git for instruction files; archives under ~/.hermes/backups/ and ~/.openclaw/backups/ (moved, never deleted).