Homeserver for mgaction.town (@user:mgaction.town). Synapse and the
Matrix Authentication Service share jupiter's postgres; logins are
OIDC-only through Authentik via MAS, since Element X can't use
Synapse's own SSO. neptun serves the apex .well-known delegation plus
matrix. and account. vhosts; the admin API stays tailnet-only.
Synapse's signing key is a sops secret so the federation identity
survives losing the RAID0 array. Covered by checks.matrix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The qBittorrent GUI runs as a system unit inside a `proton` network
namespace whose only interface is WireGuard to Proton (NL#490, P2P with
NAT-PMP) and whose only route is the tunnel, so a dropped VPN stalls
torrents instead of leaking them. DNS goes to Proton's 10.2.0.1, with
the host's nscd/resolved sockets hidden since they would answer from the
root namespace. A NAT-PMP loop keeps Proton's forwarded port mapped and
sets it through the netns-local WebUI.
qbittorrent itself is not on PATH: the qbittorrent-vpn launcher and
desktop entry (also the magnet/.torrent handler, via a watch folder)
start the unit, which polkit lets darman start and stop without a
password. The namespace comes up with the GUI and goes away with it.
The client is internet-facing, so it is sandboxed too: it sees only its
own state, the watch folder, the download dir and read-only theming;
the rest of $HOME, /mnt, darman's sops secrets, pass-broker/tailscale
sockets and the session bus are gone. The bus would let it run anything
through systemd --user, so no tray icon or notifications.
Covered by checks.qbittorrent-vpn against a fake Proton peer.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Steam is now millennium-steam from the Millennium flake's overlay (client
themes and plugins), on terra and titan. The overlay builds against our
nixpkgs; its nixpkgs-bun input stays pinned because a fixed-output hash
depends on it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Telegram bot has a single getUpdates consumer, so a broker per host
would split approval taps between them. mars now runs the gateway: the
bot's only poller, relaying prompts and taps for terra's and titan's
brokers over the tailnet (:8097) without ever seeing an item. Each broker
authenticates with its own pass_broker_gateway_token; mars holds only the
sha256, and the prompt is labelled with the host that token belongs to.
The bot token and chat id move to mars and are dropped from the other
hosts' sops files (terra.yaml also loses its leftover LibreChat keys and
gains protonvpn_wg_private_key, which a later commit uses). Covered by
checks.pass-broker-gateway (two nodes: labels, cancel, expiry, per-host
token isolation, gateway down); the fakes are shared with checks.pass-broker.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
terra's desktop is split into profiles/workstation.nix (NixOS side) and
home/desktop/ (home-manager side, moved from hosts/terra/home*), so titan,
an ASUS Zenbook Pro Duo UX581 (Intel iGPU + RTX 2060 PRIME offload),
gets the same Hyprland workstation. Per-host monitors, primary/secondary
output and wallpaper are homelab.desktop.* options; the quickshell widgets
read the outputs from QS_PRIMARY_OUTPUT/QS_SECONDARY_OUTPUT instead of
hardcoding DP-2/HDMI-A-1. Folder icons come from a new HyprChrome-Folders
icon theme.
titan gets its own sops file and recipient. LibreChat is removed from
terra along with its secrets declarations.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Cursor size 24 -> 36 (with a 36px XCursor frame so X clients don't snap
to 32). The greeter's Hyprland gets the same theme via a new
homelab.greeter.cursor option, linked into the greeter's ~/.icons.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Prompts and the journal now show "darman via Claude Code" (a `claude`
ancestor) or the terminal holding the tty, plus the caller's command with
store/profile paths cut to program names, and the process chain. It is
a hint, not identity: anything running as darman can fake it.
A client that has already hung up now always loses the race against the
prompt lock and Telegram polling, so it can no longer get a prompt, and
gone clients get no reply (no more EPIPE noise). The ghost-request test
freezes the broker to reproduce the queued-while-restarting case exactly.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A Rust daemon running as its own system user holds terra's only pass-cli
session, an agent PAT scoped to the HomeLab vaults. `pass-broker get` asks
it for an item; per-item rules decide auto/ask/deny, and "ask" sends a
Telegram prompt that only darman's account can approve (10-min grant for
darman@*/root@*). Requests whose client is gone are dropped or withdrawn,
and the client gives up after 30s without an answer.
scripts/deploy uses the broker exclusively when its socket exists. Covered
by checks.pass-broker (fake pass-cli + Telegram). Secrets added for all
hosts; only terra runs the broker so far.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
eleven_v4_turbo is the new default, same price per character as
v3_conversational. v4 also streams over Text to Dialogue, so the switch is
now uses_dialogue() for both. Audio tags may carry free-text directions
("[whispers, in a German accent]"): commas and apostrophes allowed, up to
80 chars. The voice prompt and luna-soul.md teach her the wider tag set,
including in text_to_speech voice notes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
mercury's tailscaled crash-looped on SD-card corruption and took DNS down on
every tailnet node, since pihole over the tailnet was the only global
nameserver. Cost: queries can skip pihole while mercury is slow.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Run immich and seerr as official images (their node/python stacks live in
podman's graphroot on the array), ship only the firmware the board loads,
and point jellyfin at intel-media-driver directly instead of
hardware.graphics (mesa + llvm).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
base.en heard "backup" as "Mac up" and turned German into English nonsense.
small with -l auto gets both right at ~1s per turn (vs ~0.4s); turbo was
~2.3s for no visible gain. Drops the subtitle credits multilingual whisper
invents out of silence.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011XkDJkdDzneVVdnkZfrvTY
Audio-tag and speech-stripping rules moved here from her SOUL.md: only
luna-voiced sends [voice], so they no longer ride along in every
Telegram prompt.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
pihole v6 made dns.domain an object and silently ignores a bare
FTLCONF_dns_domain, so every lease resolved as <host>.lan and names like
mars.sol (luna's apps) did not resolve at all. Set dns.domain.name.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- TTS defaults to eleven_v3_conversational over the Text to Dialogue
socket: performs audio tags and gets her German right. Flash still
works via the TTS socket.
- Every spoken message is prefixed with a [voice] marker so luna's
profile on mars can tell the channel apart.
- LunaIndicator/LunaCall replaced by a LunaPanel in the bar rail, with
captions; headless test updated to match.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
gitea-mcp in HTTP mode on :8093, reachable only over tailscale0. It
holds no token; each client sends its own PAT as a Bearer header.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The key is a toggle rather than push-to-talk, because "voice mode is
active" has to be a state something can display; a held key cannot
sustain one. So the daemon grew a VAD loop: speech starts a capture,
`silenceMs` of quiet ends it. A turn is awaited inline, which is also
what stops her hearing herself — the tick cannot fire while she speaks.
The keybind runs the CLI, not a quickshell GlobalShortcut, so the mic can
still be switched off while the shell is reloading and a QML error can
never leave it open with nothing on screen saying so.
Every threshold had to be resized for terra's mic, which is QUIET:
0.001-0.002 ambient against 0.0045 speaking, about 6dB apart. The VAD
floor of 0.006 and whisper's own 0.015 gate both sat ABOVE the user's
voice, and a 3.5x multiplier cannot separate signals that close — three
independent reasons nothing could ever trigger. Hence vadMult 1.8,
vadFloor 0.0025, and stt::normalize lifting the clip (capped at 30x, so
near-silence is not amplified into words) before whisper sees it.
The floor tracker also had to stop feeding on everything: a rising sound
dragged its own threshold up faster than it rose, and a tone peaking at
0.029 was outrun by a threshold that reached 0.018. Only samples below
the current trigger update it now.
Four diagnostics, because none of the above was visible while it was
failing: `watch` prints the socket protocol, `say` runs a turn from typed
text (a broken mic vs a broken pipeline), `level` is a live meter with
the trigger marked, `devices` lists the inputs. Amplitude is now emitted
on every tick, not only while she speaks — a meter that moves only during
playback cannot answer "is it hearing me".
Also: connect-time replay of mode/state, so restarting the shell while
voice mode is on does not leave it blind until the next event, which in
an idle conversation is never.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
luna-voiced: mic -> whisper-cpp (local, vulkan) -> Hermes' API on mars ->
ElevenLabs streaming TTS -> speaker, bound to SUPER+V. State, amplitude
and a viseme timeline are published on a unix socket; nothing consumes
that yet, but it is the data source the quickshell avatar will read.
STT stays local on purpose. mars already runs faster-whisper for Telegram
voice notes, but routing the mic through it would put every spoken word
on the network for no latency win.
Shape notes, all of them things the first live runs decided:
- Viseme offsets are relative to speech_start (emitted when the first
PCM frame reaches the speaker), never wall clock, so the shell and
the daemon never have to compare clocks. ElevenLabs sends character
alignment, not visemes, so the mapping is ours; the glTF morph target
names are the contract.
- clean_for_speech is a whitelist. She writes "(¬‿¬)", whose characters
live in Latin-1 and punctuation blocks rather than any emoji range,
so enumerating what to drop never converges. Em dashes become commas
because they are a pause — dropping one ran two clauses together and
changed what a sentence meant.
- whisper invents words out of silence ("you", "Thank you."), and a
held key over a quiet room is long AND empty, so has_speech gates on
the loudest 100ms window rather than on duration.
- Beeps on key down/up and a distinct low tone for "heard nothing":
push-to-talk with no acknowledgement is indistinguishable from broken,
and a dead mic must not feel like a slow reply.
- A warmup request at startup: her first reply after the container
starts costs ~16s against ~5s warm, and the first one is always the
one you are waiting for.
- `luna-voice watch` prints the socket protocol and `luna-voice say`
runs a turn from typed text, which is how a broken mic is told apart
from a broken pipeline.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
greetd now runs a throwaway Hyprland hosting dotfiles/quickshell/greeter.qml,
configured per host via homelab.greeter.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Comments had drifted into multi-paragraph narrative (git commit
lineage, debugging stories, restated code) in several hot spots
(scripts/deploy, hermes-agent.nix, flake.nix, gitea.nix, headscale.nix).
Trim every comment to its load-bearing "why" — gotchas, safety
warnings, and non-obvious rationale survive verbatim in substance,
just tightened to 1-2 sentences; historical narrative and anything
already covered in CLAUDE.md is cut. No code/logic changed.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01UJqEmY1y3AYX3JoX4Y6b21
Adds services/dev/obsidian-livesync.nix — CouchDB 3 from the native
nixpkgs module, tuned as the backend for the Self-hosted LiveSync plugin
— and publishes it as notes.mgaction.town through neptun.
It goes out over the public reverse proxy rather than staying on the LAN
because Obsidian's mobile apps refuse cleartext HTTP and *.jupiter.sol
cannot hold a publicly trusted cert. That makes the hardening load-bearing
rather than decorative:
- require_valid_user in both [chttpd] and [chttpd_auth], so nothing
answers unauthenticated on the open internet;
- neptun's vhost matches on CouchDB's own naming rule (system endpoints
all begin with `_`, user databases never can), so Fauxton, /_all_dbs
and /_node/_local/_config — which rewrites the server config given
admin credentials — 404 at the proxy while any number of per-vault
databases pass. Verified against both sets of paths with caddy run
against a stub backend;
- the plugin's own E2EE carries the actual confidentiality: jupiter only
ever stores ciphertext. Its passphrase is deliberately NOT in sops —
it never leaves the clients, and pairing it with the server credential
would defeat the point.
flush_interval -1 is required, not tuning: replication rides a continuous
_changes feed that caddy would otherwise buffer into a stall.
Storage sits on the array with RequiresMountsFor, since a CouchDB that
starts without /mnt/data would create an empty database on the eMMC and
LiveSync would replicate that emptiness back to every client. Logs go to
journald rather than the unrotated /var/log/couchdb.log, for the same
29G-eMMC reasons as the rest of jupiter.
The admin password reaches CouchDB as an [admins] ini fragment via
extraConfigFiles; services.couchdb.adminPass would have rendered it into
the world-readable store.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TLN5nkLBtCciD3ZnUwtw2b
The bind onto /var/lib/private/victoriametrics needs its source to exist or
the mount fails -- and because it is `nofail`, quietly: RequiresMountsFor is
satisfied by /mnt/data itself, so the service would start anyway and write
the TSDB to the eMMC, which is the one thing the bind exists to prevent.
prowlarr.nix has no tmpfiles rule only because its directory predates the
module (migrated from ZimaOS); this is a fresh service, so it creates its
own, same as seerr.nix. Verified on jupiter: the mount is live on md127 and
nothing lands on the OS disk.
scrape_timeout was left implicit at the Prometheus default of 10s, which is
longer than the 5s interval -- VictoriaMetrics clamps it down rather than
erroring, so the config claimed 10s while the scraper used 5s. Say what
actually happens. Checked with `victoria-metrics -promscrape.config.dryRun`,
not just nix eval, which never builds the checked-config derivation.
Also comments: why the bind exists and why `nofail` is load-bearing (the
fileSystems block had none, unlike prowlarr.nix and seerr.nix), and why
mercury needs its own job -- scrape_interval is per-job and job_name must be
unique, so its `job` label will always differ from the other hosts'. Select
on `host` in dashboards or mercury drops out of them silently.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
The hook registered with no events at all and delivered nothing.
"pull_request_review_comment" and "pull_request_review_rejected" are real
HookEventTypes and real X-GitHub-Event-Type values, but they are not
things gitea's hook API accepts. updateHookEvents
(routers/api/v1/utils/hook.go) matches a fixed list of api names and
silently ignores anything else, so every event flag stayed false, the POST
succeeded, and the hook sat there inert.
There is no narrower api name: HasEvent (models/webhook/webhook.go)
collapses approved, rejected and review-comment onto
HookEventPullRequestReview, so `pull_request_review` is a single switch for
all three. Approvals consequently cannot be excluded at the hook any more.
They now cross the wire as "pull_request_approved", which is not in the
route's event list, so Hermes ignores them on the event match -- before the
filter script and before any LLM call. Gitea's delivery log will show them
answered 200/ignored, which is intended.
That makes three namespaces for the same event rather than two, so the
tables in both nix files and the README now carry the api column, and the
README warns about the silent-ignore behaviour that hid this.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
One webhook per route, from a list, so adding a route is an entry rather
than a copy of the unit. The PR-review hook subscribes
pull_request_review_comment and pull_request_review_rejected.
The unit runs as the gitea user on a multi-user box, where
/proc/<pid>/cmdline is world-readable for the lifetime of the process, so
`-H "Authorization: token $t"` published the admin token and
`jq --arg secret "$s"` the webhook secret -- which is exactly what the
existing comment claimed to be avoiding by putting the body on stdin. The
token now goes through a 0600 `curl -K` config written with printf (a
shell builtin, so the substitution never reaches an argv) and the secret
through jq --rawfile.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
A timeline comment on a PR never reached the route. Gitea reuses the same
strings in two namespaces and they collide:
subscription name wire name (X-GitHub-Event) what it is
pull_request_comment issue_comment comment on a PR
issue_comment issue_comment comment on an issue
pull_request_review_comment pull_request_comment review on a PR
The hook's `events` array takes the subscription name; Hermes matches
--events against X-GitHub-Event, the wire name, produced by
HookEventType.Event() in modules/webhook/type.go. So --events
pull_request_comment was selecting review submissions and could never match a
comment -- the exact inversion of what it reads like.
That also explains both observed failures. The review submission matched
(wire name pull_request_comment) and reached the filter, which correctly
dropped it on action=reviewed since a PullRequestPayload carries no comment
object. The timeline comment arrived as issue_comment, matched nothing, and
was dropped by the events filter before the script ever ran.
gitea.nix and hermes-agent.nix now deliberately name the same event
differently, so both carry the table and say the other is not a typo.
issue_comment on the wire also covers comments on plain issues. The hook does
not subscribe those, and the filter's is_pull check drops them regardless, so
widening the hook later cannot leak issue comments into the agent.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
Names the webhook "PR comments Hermes" (gitea's CreateHookOption/EditHookOption
both carry an optional `name`, so it survives the create and the update path)
and narrows it from all 26 event types to pull_request_comment alone.
Gitea sends pull_request_comment distinctly from issue_comment, so the hook
now covers comments on pull requests and nothing else. Hermes would have
dropped the rest anyway -- its route filters on X-GitHub-Event before any LLM
call -- so this is defence in depth rather than the only gate, but it keeps
traffic that can never be acted on from crossing the wire and reaching the
agent's process at all.
The tradeoff is that event selection now lives on both sides: a second Hermes
route needs its event adding here as well as being subscribed. That is the
right way round for a single-purpose hook, and the comment says so.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
The relay existed on the premise that Gitea sends no header Hermes can read
an event name from, so something had to copy X-Gitea-Event into
X-GitHub-Event. That premise was wrong. Gitea's addDefaultHeaders sets
req.Header["X-GitHub-Delivery"] = []string{t.UUID}
req.Header["X-GitHub-Event"] = []string{event}
req.Header["X-GitHub-Event-Type"] = []string{eventType}
unconditionally, for every webhook type, alongside X-Hub-Signature-256 in
GitHub's exact format. (Direct map assignment rather than .Add() specifically
to keep the "GitHub" casing that canonicalisation would destroy.) Hermes
validates that signature on any route without provider gating and reads the
event name from that header, so gitea and hermes already speak the same
protocol and the translation layer was translating nothing.
Gitea now posts straight at http://mars.orbit.sol:8644/webhooks/gitea-pr-comments.
The URL path is the Hermes route name, so a second subscription is a second
hook and nothing else -- the route-in-path indirection the relay grew was a
reimplementation of something Hermes already had.
Removes the module, the 200-line relay, its test, the mars import, the 8645
listener, and the stale gitea-hermes-webhook-relay.service entry left in the
secret's restartUnits. hermes-agent-webhook-route moves to
hosts/mars/hermes-agent.nix, next to the container and the read-only prompt
and filter mounts it depends on.
Also makes that unit refuse to subscribe when GITEA_HERMES_WEBHOOK_SECRET is
unset in the container, matching the existing empty-prompt check. An empty
secret silently fails every delivery signature check afterwards while the
unit still reports success -- the worst possible failure shape, and one this
setup can actually produce on a first deploy.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
Mirrors the `hermes` alias on mars. The admin CLI is effectively
undiscoverable without it: the package is not in systemPackages so `gitea` is
not on PATH at all, every admin subcommand needs GITEA_WORK_DIR pointed at a
stateDir that is not the module default, and it has to run as the gitea user
or it drops root-owned files into that directory. Getting any of the three
wrong fails in a different and unhelpful way.
Both the package path and the stateDir come from the config rather than being
written out, so a gitea bump or a stateDir move cannot leave the alias
pointing at something stale — which is exactly what a hardcoded /nix/store
path would do.
Lives in services/dev/gitea.nix, which only jupiter imports, so it does not
leak onto hosts with no gitea to administer.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
Webhook delivery to the hermes relay was refused outright:
Post "http://mars.orbit.sol:8645/gitea/gitea-pr-comments":
dial tcp 100.64.0.6:8645: webhook can only call allowed HTTP servers
(check your security.ALLOWED_HOST_LIST setting),
deny 'mars.orbit.sol(100.64.0.6:8645)'
ALLOWED_HOST_LIST defaults to `external`, documented as "a valid non-private
unicast IP". Tailscale addresses come from 100.64.0.0/10 — RFC 6598
carrier-grade NAT space — which is not RFC1918 private but does not satisfy
gitea's notion of external either, so every tailnet target is denied by
default. Nothing about the relay or the URL was wrong; the request never left
jupiter.
Sets the tailnet CIDR explicitly and keeps `external`, so a future webhook to
a public service still works without another edit here.
Goes in [security], not [webhook]: the webhook-section key is deprecated in
favour of this one and now merely falls back to it, and [security] is the
name the delivery error itself reports.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
Retiring gitea-events is a one-off migration, not something worth re-running
on every boot. Both units now only touch what they own: jupiter's creates or
updates its own hook and deletes nothing, and mars's removes only the route
it is about to re-subscribe, as the idempotency step for `subscribe`.
Keeping the deletes would have meant a redeploy could silently remove a hook
or route someone added deliberately -- a real risk now that sibling hooks
for other Hermes routes are the intended pattern.
README carries the manual commands, and the note that both hooks fire until
the old one is removed by hand, so events arrive twice in the meantime.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
Renames the subscription to gitea-pr-comments (it handles one event; the old
gitea-events name promised more than it delivered) and drops --deliver.
Rather than move the hardcoded route from one constant to another, the relay
now reads it from the request path: POST /gitea/<route> forwards to
<base>/webhooks/<route>. The route name was the last thing tying this service
to a specific subscription, so a second Hermes route is now a `hermes webhook
subscribe <name>` plus a Gitea hook at /gitea/<name>, with no relay change --
previously it would also have needed a second relay URL baked in here.
The path segment is interpolated into an outbound URL, so it is validated
against ^[A-Za-z0-9][A-Za-z0-9._-]{0,63}$ and refused rather than sanitised
when it does not match. The path is matched raw and never URL-decoded, so
percent-encoded separators fail the charset check instead of surviving it;
requiring an alphanumeric first character also rejects "." and "..". Without
this, POST /gitea/..%2fadmin would let anything that can reach the relay
steer it at other Hermes endpoints. Tests cover traversal, encoded traversal,
embedded slashes, leading dot/dash, and the length bound, and assert nothing
reaches the stub Hermes in any of those cases.
Dropping --deliver leaves it at its default of `log`. The prompt tells her to
answer in the pull request, so the PR comment is the delivery and a Telegram
copy would only duplicate it; this also removes the hardcoded chat id that
was a third copy of TELEGRAM_HOME_CHANNEL.
Provisioning retires the pre-rename hook by its EXACT old URL rather than by
"points at the relay". Now that sibling hooks for other routes are the
intended pattern, a prefix match would delete them.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
Completes the subscription: it had a secret, a delivery target and a script,
but no prompt and no event list, so it woke the agent on every forwarded
event with nothing to tell her what to do.
--events pull_request_comment narrows the route to the one event the prompt
handles. This only works because the relay copies X-Gitea-Event into
X-GitHub-Event; without that every delivery arrives as "unknown" and matches
nothing. Gitea sends pull_request_comment distinctly from issue_comment, so
plain issue comments no longer reach the agent at all. The Gitea-side hook
still posts the full event set to the relay and Hermes drops the rest before
any LLM call.
The prompt lives in hosts/mars/gitea-pr-comment-prompt.md, mounted read-only
next to the filter, and is read with $(cat) at subscribe time rather than
passed inline. That is not only about escaping: the text has to survive nix
`` string escaping, the systemd unit file, and `podman exec sh -c '...'`
single-quoting. It contains an apostrophe ("the PR's head branch") that
would terminate that single-quoted string early. Read from a file at runtime
the content never passes through shell source, so it can contain anything.
Verified end to end against the rendered unit with stubbed podman/hermes:
the value reaching --prompt is byte-identical to the repo file apart from
the trailing newline that command substitution strips.
`set -eu` inside the container shell is load-bearing. Without it a missing
prompt file makes cat fail, the substitution yields "", and the subscription
is created with an empty prompt -- a silent failure that still looks like a
healthy unit.
On what read-only does not buy: it protects the sources, and this unit
re-subscribes from them on every start, so a restart restores the intended
prompt, filter and events. The live subscription itself lives in
webhook_subscriptions.json under /opt/data and is hot-reloaded, which is
inside the agent's own write-safe root -- a self-modification would stick
until this unit next runs.
The prompt keeps its own stop conditions even though the filter already drops
those deliveries, and says explicitly that reaching them means the filter
failed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
The gitea-events subscription woke the agent on every delivery. That is an
unbounded loop as soon as she is given a prompt that tells her to answer on
the PR: her answer is itself a pull_request_comment, which wakes her again.
Adds a Hermes route script that drops the deliveries that must never reach an
LLM call: luna's own comments (the loop guard), "deleted" actions (the body
is still in the payload, so acting on one means acting on a request that was
explicitly withdrawn), non-pull-request comments, empty bodies, and edits
that did not actually change the body — a label or attachment change fires
"edited" too. Everything else passes through unchanged.
Mounted READ-ONLY from the nix store rather than written into hermesHome.
Hermes resolves route scripts under ~/.hermes/scripts, which here is inside
/opt/data — HERMES_WRITE_SAFE_ROOT — so a filter written there would be a
loop guard sitting in the writable root of the agent it constrains. Deleting
it fails closed (Hermes treats a missing script as "ignore"), but rewriting
it to always-allow would silently restore the loop. Read-only from the store
makes that impossible and keeps the guard in git.
The script also normalises changes.body.from to always exist. Gitea omits
`changes` entirely on created events, and Hermes replaces the prompt payload
with whatever JSON the script emits, so guaranteeing the key here means a
prompt referencing {changes.body.from} renders empty instead of leaving an
unfilled placeholder.
Note the stdout contract (gateway/platforms/webhook.py): only exactly
"[SILENT]", empty output, or a nonzero exit drop a delivery. Any OTHER text
on stdout lets it through and is attached as script_output — so a stray
debug print would silently defeat the filter. All diagnostics go to stderr,
and gitea-pr-comment-filter-test.py asserts that discipline along with each
drop rule (25 cases). Run it after any edit: the fail-closed behaviour means
a syntax error produces silence, not an error.
--events is still unset; event selection remains runtime-tunable policy.
The filter covers only what must not be.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa
Three fixes to how luna's git/tea credentials are set up on mars, all found
against the running instance on 2026-08-23.
Drop the host-side clone. hermes-agent-prepare-dirs used to clone this repo
into ${hermesHome}/workspace/homelab, but nothing ever told luna at runtime
that it was there — she self-manages config/profiles/memories, so a path
baked into this file never reached her. She searched /opt/data/homelab and
/workspace, found neither, and concluded she had no repo at all. The
credentials are what actually grant access; any checkout is hers to make
anywhere inside HERMES_WRITE_SAFE_ROOT. The stale directory left by the old
version is deliberately not cleaned up, just unmanaged from here on.
Point credential.helper at the CONTAINER's path. It was written as the host
path (${hermesHome}/.git-credentials), which does not exist inside the
container where git actually reads the config — broken this way from 3c1f3e5
until now. Nothing host-side consumes those credentials any more, so the
container's view is the only one that has to be right; added `containerHome`
to make the distinction explicit at the point of use.
Chown what the oneshot writes. The image's cont-init only chowns the top
level of hermesHome and its own state — it does not recurse into the
root-owned 0600 files this unit drops there (.git-credentials, and tea's
config.yml, which tea also writes 0600), even though it runs afterwards. The
symptom was not an error but an absence: git reported no credential helper
and tea no login. Uses `if` rather than `[ -d x ] && chown` because under
`set -e` a false test on the left of an && list aborts the unit.
gitea.nix carries the matching comment updates: the luna provisioning unit is
server-side only, and her token needs write:repository,write:issue,read:user.
write:issue is the one that is easy to miss — a pull request IS an issue in
gitea's data model, so /pulls endpoints gate on the issue scope category and
`tea pr create` fails with write:repository alone even though clone, fetch
and push all work.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S94o42aQ8VkBmEWvDem5xa