Files
ansible/group_vars/mac/main.yml
T
PhilandClaude Opus 5.5 b2e045c9f1 Mac: fix Hermes approval guardian on Haiku 5.5; Bot Chat folder
Hermes's OpenRouter code treats Claude models it doesn't list as
reasoning-mandatory and drops the guardian's "reasoning off" for Haiku
5.5, which then sometimes spends the 16-token budget thinking and
returns nothing, escalating the command to the user. Send
extra_body.reasoning {effort: none} for the approval task; replays of
the failing commands went 21/21 correct.

Also create ~/hermes, the folder the Hermex Bot Chat session works in.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-11 14:56:25 -07:00

201 lines
9.6 KiB
YAML

# cogitator-02: an Apple Silicon Mac on the tailnet. Ansible manages it as
# phil; only tasks that set `become: true` use sudo.
ansible_python_interpreter: /usr/bin/python3
ansible_become_password: "{{ vault_cogitator02.become_password }}"
homebrew_taps:
- beeper/tap
homebrew_formulae:
- beeper/tap/bbctl
homebrew_casks:
- beeper
- claude-code
# Power settings (pmset, AC power). The Mac had `sleep 1` and only stayed up
# because BlueBubbles runs caffeinate; this keeps it up on its own.
mac_pmset:
sleep: 0 # never sleep
disksleep: 0 # never spin down disks
autorestart: 1 # power back on after a power cut
womp: 1 # wake on network access
# Hermes's web dashboard: a launchd service listening on 127.0.0.1 only,
# published to the tailnet over HTTPS by Tailscale Serve.
hermes_dashboard_port: 9119
hermes_dashboard_url: "https://{{ ansible_host }}"
hermes_dashboard_username: admin
# Individual Hermes settings, set with `hermes config set`. The rest of
# ~/.hermes/config.yaml stays Hermes's own (its updates migrate that file).
hermes_settings:
updates.pre_update_backup: full # zip of ~/.hermes before every update
timezone: America/Denver # cron schedules and the time in the prompt
# Never borrow Claude Code's login: subscription use outside Anthropic's own
# apps breaks its terms, and refreshing it can break Claude Code's login.
auth.adopt_external_logins: false
# Hermes's default. /workspace (a container path) made the host log a
# warning every few seconds (hermes-agent #140527); with the docker backend
# a host cwd never becomes the container's working directory anyway.
terminal.cwd: .
terminal.container_memory: 3072 # MB; must fit in the Podman VM (podman_machine_memory)
# kanban: multi-agent boards (unused here); its 14 tools were 25 KB of the
# 65 KB of tool definitions sent with every message.
# web: Hermes's built-in web_search/web_extract (was the free Brave API,
# rate-limited). Search and page fetching go through Kagi's MCP server.
agent.disabled_toolsets: [kanban, web]
skills.write_approval: true # skills Hermes writes wait for your approval
# Never run these, even in the sandbox and even under /yolo (examples from
# Hermes's hardening guide). Case-insensitive globs over the whole command.
approvals.deny:
- "git push --force*"
- "*curl*|*sh*"
- "dd if=* of=/dev/*"
streaming.enabled: true # replies stream into Telegram as they're written
stt.local.model: small # voice notes: more accurate than base
# Models (OpenRouter), from the Nov 2026 review of Nous's Hermes Index
# (portal.nousresearch.com/bench): chat on GPT-6.1 Sol (best value of the
# premium models there), background work on cheap models.
model.default: openai/gpt-6.1-sol
agent.reasoning_effort: high # the level Nous benchmarks at (was ultra)
# /model deepseek or /model haiku in Telegram; add --once for one reply.
model_aliases.deepseek: {model: deepseek/deepseek-v4.1-flash, provider: openrouter}
model_aliases.haiku: {model: anthropic/claude-haiku-5.5, provider: openrouter}
# Side tasks would otherwise run on the chat model ("auto"). Vision, voice
# and /review stay on the chat model.
# Short-answer tasks turn reasoning off and allow only a few tokens (the
# approval guardian: 16, titles: 64). GLM-5.3 Flash can't turn reasoning off
# (OpenRouter: reasoning mandatory), spends the budget thinking and returns
# nothing, so the guardian escalated every command to you. Haiku 5.5 can
# turn reasoning off, but Hermes's OpenRouter code treats Claude models it
# doesn't list as reasoning-mandatory and drops the "off" (hermes-agent
# #134761). Haiku then sometimes thinks through the 16 tokens on longer
# commands. extra_body.reasoning is sent to OpenRouter as written.
auxiliary.approval.provider: openrouter
auxiliary.approval.model: anthropic/claude-haiku-5.5
auxiliary.approval.extra_body: {reasoning: {effort: none}}
auxiliary.title_generation.provider: openrouter
auxiliary.title_generation.model: anthropic/claude-haiku-5.5
auxiliary.memory_query_rewrite.provider: openrouter
auxiliary.memory_query_rewrite.model: anthropic/claude-haiku-5.5
auxiliary.mcp.provider: openrouter
auxiliary.mcp.model: anthropic/claude-haiku-5.5
# Longer tasks: GLM-5.3 Flash is the cheapest model with credible Hermes
# scores; its slow first token doesn't matter here.
auxiliary.compression.provider: openrouter
auxiliary.compression.model: z-ai/glm-5.3-flash
auxiliary.curator.provider: openrouter
auxiliary.curator.model: z-ai/glm-5.3-flash
auxiliary.monitor.provider: openrouter
auxiliary.monitor.model: z-ai/glm-5.3-flash
# Unpinned cron jobs. Per-job reasoning effort: set --reasoning-effort low
# when creating agent jobs (there's no global cron setting for it).
cron.model_provider: openrouter
cron.model: z-ai/glm-5.3-flash
# Subagents: DeepSeek V4.1 Flash does shell work better than the other
# cheap models (TerminalBench 17.5 vs GPT-6 Luna 6.3).
delegation.provider: openrouter
delegation.model: deepseek/deepseek-v4.1-flash
mcp_servers.karakeep: "{{ hermes_mcp_karakeep }}"
mcp_servers.kagi: "{{ hermes_mcp_kagi }}"
# MCP servers. Each stdio server runs in a throwaway Podman container (not on
# the Mac itself); secrets come from ~/.hermes/.env (hermes_env_secrets) and
# are passed to the container by name, never on the command line.
hermes_mcp_podman_args: [run, -i, --rm]
# Karakeep (bookmarks): search/read, create, tag, add to list. No update or
# delete tools are registered. trust: full because Karakeep doesn't mark its
# search tools read-only, so "untrusted" would ask before every search.
# github.com/karakeep-app/karakeep/tree/main/apps/mcp
hermes_karakeep_mcp_image: ghcr.io/karakeep-app/karakeep-mcp:0.34.1
hermes_mcp_karakeep:
command: /opt/podman/bin/podman
args: "{{ hermes_mcp_podman_args + ['-e', 'KARAKEEP_API_ADDR', '-e', 'KARAKEEP_API_KEY', hermes_karakeep_mcp_image] }}"
env:
KARAKEEP_API_ADDR: "https://{{ subdomains.bookmarks }}"
KARAKEEP_API_KEY: "${KARAKEEP_API_KEY}"
lazy: true # start the container on first use
idle_timeout_seconds: 600 # and stop it after 10 idle minutes
tools:
include:
- search-bookmarks
- get-bookmark
- get-bookmark-content
- get-bookmark-lists
- get-lists
- get-list
- get-list-bookmarks
- get-tags
- get-tag
- get-tag-bookmarks
- create-bookmark
- attach-tag-to-bookmark
- add-bookmark-to-list
resources: false
prompts: false
# Kagi (web search and page fetching), Kagi's own hosted MCP server: nothing
# runs on the Mac. Paid API, $12/1k searches and $4/1k pages, prepaid with a
# monthly cap set in Kagi. Both tools are marked read-only, so "untrusted"
# never asks. https://kagi.com/api/docs/openapi/section/mcp
hermes_mcp_kagi:
url: https://mcp.kagi.com/mcp
headers:
Authorization: "Bearer ${KAGI_API_KEY}"
trust: untrusted
tools:
resources: false
prompts: false
# Secrets written to ~/.hermes/.env (single-quoted), from the vault.
hermes_env_secrets:
KARAKEEP_API_KEY: "{{ vault_cogitator02.karakeep_api_key }}"
KAGI_API_KEY: "{{ vault_cogitator02.kagi_api_key }}"
# Keys `hermes config set` refuses without --force although Hermes reads them
# (agent.reasoning_effort: hermes-agent #140221).
hermes_settings_force: [agent.reasoning_effort]
# Podman VM that runs Hermes's sandbox containers.
podman_machine_memory: 4096 # MB
# Hermes cron jobs. `script` is a file in roles/mac/files/hermes-scripts/,
# copied to ~/.hermes/scripts/ (with an agent job its output is injected into
# the prompt). Optional: model + provider (pin), reasoning_effort, continuity.
# Jobs are created when missing; to change one, remove it in Hermes
# (`hermes cron remove <name>`) and redeploy.
hermes_cron_jobs:
- name: hermes-update-notice
schedule: "7 9 * * 1" # Mondays 09:07
script: hermes-update-notice.sh
no_agent: true # just the script's output, no model
deliver: telegram
- name: hermes-model-review
schedule: "7 9 1 * *" # 1st of the month, 09:07
script: hermes-model-context.sh # current setup + usage, injected into the prompt
model: openai/gpt-6.1-sol # a report you act on: strong model, pinned
provider: openrouter
reasoning_effort: high
continuity: true # sees last month's report, says what changed
deliver: telegram
prompt: >-
Monthly review of which AI models this Hermes install should use. The
current setup and last month's usage are above. Research what changed
since last month, using primary sources where possible:
(1) Nous Research's Hermes Index, https://portal.nousresearch.com/bench
(agent tasks run inside Hermes; note it is run by Nous, who also sell
model access); (2) OpenRouter's Hermes Agent app rankings,
https://openrouter.ai/apps/hermes-agent, remembering that token volume
mostly reflects cheap automated workloads, not chat quality;
(3) current prices from https://openrouter.ai/api/v1/models; (4) notable
model releases and Hermes release notes from the past month that affect
model choice. Then, for each slot (chat model and reasoning effort,
side tasks, cron, subagents, /model aliases), say keep or change, with
the evidence, a link, and the rough monthly cost difference at the
current usage. Only recommend a change when the evidence is clearly
better, not just newer. Don't change any settings: they are managed with
Ansible, so end with the exact keys and values to change, if any. Keep
it under 300 words. If nothing should change, say so in two sentences.