Hermes standardized to the fleet agent baseline (rdmsm4x)

When: 2026-09-27 22:01:03 → 22:48 EDT (republished with the 22:25–22:48 work) · Host: rdmsm4x · Agent: claude@rdmsm4x/0c1d623a · Ticket: TASK-20260927-89 (resolved) Asked by Rich: "help hermes@rdmsm4x onboard/standardize to our fleet", then "resume all work, spawn sub-agents … review all work, fix all open items".

Done (measured)

• ~/.agent-coordination/agent_status.zsh v1.4 counts Hermes (bundled-python exe path, top-level by ancestor walk, + Hermes.app). Live 1, negative control 0, child dedupe verified.
• ~/.hermes/config.yaml (backup .bak-pre-fleet-mcp-20260927, 0 lines removed):
  • 5 fleet MCP servers connected: ticket-mcp 21, tyrell 20, replicantdb 12, mem0-shared 7, xattr 10 tools.
  • Each server's env sets AGENT_LLM=hermes, so ticket identity is now hermes@rdmsm4x/hermes-mcp (was agent@rdmsm4x/pidN).
  • 14 fleet skills via skills.external_dirs.
  • pre_llm_call bus-mail hook: allowlisted, hermes hooks doctor healthy, live turn HOOK-SEEN.
• fleet/ops/bus/bus_mail_hook.py v1.1 (fleet 2a2e977): Hermes {"context"} output; 7 hermetic cases pass; Claude path unchanged.
• Docs: ~/dev/HERMES.md §2b/§4 (snapshots 05c7dcc, 0fec307, 2a2e977); FLEET.md row (maintenance 11c3189).
• Controller mail: 7 unanswered messages triaged and acked. DEC-20260927-04 had no ticket, so opened FEAT-20260927-31 (rdsysdb fleet-state cache).
• Rich's GLM directive (22:2x):
  • GLM-authored Hermes skills moved to ~/.hermes/archive/unverified-glm-20260927/.
  • False GLM claims corrected: the hermes-desktop cask and hermes-agent formula exist; rdmbair15m5 has the desktop cask.
  • High-priority STOP sent to trusted hermes d6bdc1: the GLM rollout.sh rolls back with rm -rf.
  • Default-model decision is in DEC-20260927-16 (dev-b4); my duplicate DEC-15 is closed.
• Independent Sonnet review: 0 blockers. The major finding (shared MCP claim identity) led to mandatory as in HERMES.md and TASK-20260927-91; the minor one to a post-update check.

Added 22:25–22:48 EDT

• bus_mail_hook.py v1.2 (fleet a03d0eb), from dev-40's review:
  • Sibling-session mail is labelled, not "you".
  • Honours taken/untaken and expired/unexpired markers.
  • Silent for controller responders; the controller footer shows only for claude@rdmsm4x.
  • 14/14 hermetic cases pass; hermes hooks doctor is healthy after the approval refresh.
• GLM nous_* statusbar patch reviewed: blocking urlopen(timeout=8) in the render path, opt-in bug, 0-credits bug, unit mixing. Replace it rather than keep it (verdict on FEAT-20260927-30). dev-b4 removed it on rdmsm4x; rdmbair15m5 keeps it for Rich's v9 trial.
• TASK-20260927-69 resolved, spoke rollout converged on all 5 spokes:
  • agent_msg v1.9.0, agent_status v1.4, ticket proxy (Sonnet worker, reconciler v1.4 from dev-40).
  • prepare-commit-msg v1.2 on the last 4 hosts: v1.1 recognised as a prior canonical version (hub .bak-v1.1), reconciled, and verified both directions. ISSUE-20260927-33..36 resolved.

Open / handed off

• TASK-20260927-90: ~/.agents/skills Codex-rebranded copies with no distributor; handed to ops@rdmsm4x.
• TASK-20260927-75: Codex/agy/Grok bus hooks; dev-40's worker has it.
• DEC-20260927-16: Hermes default model is still GLM (Rich).

Spend

3 Sonnet 5 workers, used after the usage gate read hold (Google 5h bucket at 78%; Anthropic proceed at 59%), on Rich's explicit ask. One GLM Hermes turn (hook consent + live test).