Migration 023 adds notes.title/summary (no prod notes start with a heading,
so no files are rewritten). Backup: /data/backup-pre-v2.16.0.db and
/data/backup-pre-v2.16.0-files.tgz in the starpunk-data volume.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add the starpunk-mcp container (same image, uvicorn :8001, shared env via
anchor, BACKGROUND_WORKERS_ENABLED=false, socket healthcheck) and route
/mcp and /.well-known/oauth-protected-resource to it in Caddy. Bump
starpunk_version to v2.15.0 (needs the v2.15.0 tag built first).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Podman reports the sandbox image as docker.io/nousresearch/...; Hermes
compares image names as text, so the short name made it delete and
recreate the sandbox on every reuse. Use the full name.
Hermes 0.23.1 looks for Chromium in chrome-linux/, but the ARM64 image
ships chrome-linux-arm64/ (hermes-agent #153626), so browser tools said
there was no Chromium. Mount a folder with a chrome-linux/chrome script
that runs the image's Chromium. Tested in a fresh sandbox: the probe
finds it and browser_exec loads pages.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
terminal.cwd "." resolved against each process's working directory, so
the dashboard's sandbox mounted / (the whole disk, writable) and the
gateway's mounted ~/.hermes (including .env). Point it at ~/hermes.
The cua-driver symlink in ~/.hermes/skills made Hermes give each sandbox
a temp copy of the skills and delete the previous copy, emptying running
sandboxes' skills (hermes-agent #53630). Read Cua's pack through
skills.external_dirs and remove the symlink.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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>
Hermes's built-in web search was the free Brave API, which returns 429s
under agent bursts; real-user reports (Hermes #80201, PR #36838, HN) also
rule out self-hosted SearXNG for always-on use. The user has Kagi, so:
- Kagi's own hosted MCP server (https://mcp.kagi.com/mcp, Bearer API key,
paid API with a monthly cap set in Kagi). Its two tools,
kagi_search_fetch and kagi_extract, are marked read-only, so
trust: untrusted never asks.
- agent.disabled_toolsets gains `web`, removing the built-in
web_search/web_extract everywhere, so Kagi is the only search path.
- KAGI_API_KEY from the vault into ~/.hermes/.env.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The smart-approval guardian asks for a one-word verdict with reasoning off
and max_tokens=16 (title generation: 64). GLM-5.3 Flash's reasoning is
mandatory on OpenRouter, so Hermes floors it to effort=low, the model
spends the budget thinking and returns nothing (finish_reason=length), and
every flagged command escalated to the user. Move approval, titles, memory
query rewrite and MCP sampling to Claude Haiku 5.5 (reasoning off by
default); compression, curator and the mail monitor stay on GLM.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Hermes can search and read bookmarks, create them, tag them and add them
to lists from Telegram. Per Karakeep's MCP README and Hermes's MCP docs:
- the official server image (pinned 0.34.1) runs in a throwaway Podman
container (`podman run -i --rm`), started on first use and stopped
after 10 idle minutes;
- only search/read/create/tag/add-to-list tools are registered (no update
or delete); trust: full because Karakeep doesn't mark its search tools
read-only;
- the API key is in the vault, written to ~/.hermes/.env via the new
hermes_env_secrets, and passed into the container by name.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Models, from the Nov 2026 review (Nous Hermes Index, OpenRouter data):
- chat: openai/gpt-6.1-sol (user's pick), reasoning high (was ultra)
- side tasks (compression, approval, titles, mcp, memory query rewrite,
curator, mail monitor) and unpinned cron: z-ai/glm-5.3-flash
- subagents: deepseek/deepseek-v4.1-flash
- /model aliases: deepseek, haiku
agent.reasoning_effort needs --force (hermes-agent #140221), so settings
can list keys in hermes_settings_force. `config get` exits 1 for unset
keys without a default; that now reads as unset, not a failure.
hermes-model-review: monthly cron job on Sol (pinned, high effort,
continuity) that redoes this research and sends keep/change
recommendations as Ansible keys to Telegram. A script injects the live
model setup and 30-day usage into its prompt. hermes_cron_jobs gains
model/provider, reasoning_effort and continuity.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Non-interactively, `hermes config get` prints only the default for a key
that isn't in config.yaml, so an unset key can't be detected and the play
kept trying (and failing) to unset terminal.cwd. Set it to Hermes's own
default "." instead and drop the unset branch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Via hermes_settings (now compared structurally, with null = unset, and a
gateway restart on change):
- timezone America/Denver (was unset)
- auth.adopt_external_logins false: never borrow Claude Code's login
- terminal.cwd unset: stops the /workspace warnings (#140527)
- agent.disabled_toolsets [kanban]: unused, 25 KB of tool schema per call
- skills.write_approval true
- approvals.deny: force-push, curl|sh, dd to devices (docs' examples)
- streaming.enabled true, stt.local.model small
- terminal.container_memory 3072, inside a 4 GB Podman VM (was 2 GB VM
with a 5 GB container limit), set by a new podman task
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Hermes has no auto-update for itself, and /update from Telegram can leave
a launchd gateway offline. Instead, a no-agent Hermes cron job runs
`hermes update --check` on Mondays at 09:07 and messages Telegram only when
a new stable release is out (a failing check sends an error alert).
Adds hermes_cron_jobs (scripts in files/hermes-scripts/, copied to
~/.hermes/scripts/; jobs created when missing) so the deal checker can
follow the same pattern.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds a small hermes_settings map that Ansible checks with
`hermes config get` and fixes with `hermes config set`, leaving the rest
of config.yaml to Hermes (its updates migrate that file).
First setting: updates.pre_update_backup: full (was false), so every
`hermes update` zips ~/.hermes first. Rollback isn't supported once a new
version has written state.db, so the backup is the way back.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The dashboard used to be started by hand as `serve --host 0.0.0.0`, so it
was reachable from the LAN and didn't survive a reboot or a Hermes update.
Now, following Hermes's docs:
- launchd agent local.hermes-dashboard runs `hermes dashboard` on
127.0.0.1:9119 (KeepAlive, --skip-build since `hermes update` builds it).
- Tailscale Serve publishes it on https://cogitator-02.porgy-porgy.ts.net,
tailnet only. No module exists for Serve; the task reads
`serve status --json` and only sets it when it differs.
- Login: scrypt password hash and a real signing secret from the vault
(the old secret was the literal text "$(openssl rand -base64 32)"); the
plaintext password line is removed from .env since it overrides the hash.
- After Serve is set, the play checks /api/status demands the basic login;
if not, it turns Serve off and fails.
- Hermes's Python no longer needs a firewall exception, so that's gone.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- pmset: never sleep, no disk sleep, restart after a power cut, wake on
network. The Mac had sleep 1 and only stayed up because BlueBubbles runs
caffeinate.
- Application firewall on, keeping signed software allowed automatically,
stealth and block-all off. Hermes's ad hoc signed Python (dashboard on
9119) is allowed explicitly, since an unattended Mac blocks unsigned
programs silently.
- Local id_ed25519.pub in phil's authorized_keys (already there).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
cogitator-02 is an Apple Silicon Mac on the tailnet running Hermes Agent.
It gets a `mac` group and mac.yml, kept apart from site.yml so it never
runs with the server's deploys (PLAYBOOK=mac.yml bin/deploy).
The mac role starts by recording the Homebrew taps, formulae and casks
already installed (present only, never upgrades or removes), so the first
run changes nothing.
No passwordless sudo on the Mac: Hermes runs as phil and takes commands
over Telegram. Ansible's sudo password is vault_cogitator02.become_password
and only tasks with become: true use it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The 1.7 -> 1.22 upgrade is done; the app gets the dockcheck label.
Redis and Postgres stay excluded in dockcheck.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1.17.0 background jobs drained cleanly. 1.18.1 (skipping 1.18.0, which
needed the hotfix) converts Areas to Places and re-detects visits across
the whole history; the changes can't be rolled back, so a fresh dump was
taken first (~/backups/dawarich-1.17.0-20261109). Redis goes to 8 as in
upstream's compose since 1.18.0.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Follows the 1.17.0 upgrade notes and upstream compose: the separate
dawarich_sidekiq service is gone (removed on the host before deploying),
the app runs Sidekiq embedded (SIDEKIQ_EMBEDDED=true, WEB_CONCURRENCY=0,
BACKGROUND_PROCESSING_CONCURRENCY=2), gets stop_grace_period 60s, and the
limits match upstream (1 CPU, 4G).
Fresh dump before this step: ~/backups/dawarich-1.16.0-20261109.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1.14.5 background jobs fully drained (checked queues, scheduled and retry
sets, and job logs). 1.16.0 is the last release with the separate sidekiq
container and merges duplicate legacy points.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Both earlier failures had one cause: the host still had old images cached
under the :latest tag (paperless-ngx from 2025-09, i.e. 2.x; heyform from
2025-02, i.e. v2), and the compose task only pulls missing images. So the
switch to :latest started year-old code. Pulled the real :latest images on
the host. Paperless 3.4.2 has no pending migrations on this database
(checked on a copy), and HeyForm v3.0.6 listens on 9157 as before.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
1.10.3 migrated cleanly and its background jobs have drained. 1.14.5
fixes the registration bypass (GHSA-x483-3xv3-vp52) and runs the
1.11-1.14 track/visit recalculation and point backfill.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- HeyForm v3.0.6 changed the default listen port from 9157 to 8000, so
Caddy got 502s. Set APP_LISTEN_PORT=9157 explicitly.
- Paperless 3.4.2 fails its migration check on this database
(InconsistentMigrationHistory, squashed 1022..1036). Rolled back to 3.4.1
until that is investigated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
First stop is 1.10.3 (drops the legacy lat/lon columns with the 1.10.2
lock fix). Next stops: 1.14.5 (security fix), 1.16.0, 1.17.0 (sidekiq
merged into the app), 1.18.1, 1.22.0, then back to latest.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Pinned versions never got updates, because dockcheck only notices a new
image under the same tag.
- paperless-ngx: 3.4.1 -> latest (stable only; beta/dev have own tags)
- paperless redis: 7 -> 8-alpine (reads Redis 7 data as-is)
- gotenberg: 8.39.0 -> 8 (9 changed the API Paperless calls)
- tika: 3.3.1.0 -> 3.3.3.0 (no floating 3.x tag; Tika 4 breaks Paperless)
- heyform: v3.0.5 -> latest (only moved by v* release tags)
- add mag37.dockcheck.update to all of the above
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>