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>
tome dropped out of home.packages in an auto-stash merge (6406e06). Its
NuGet lockfile also pinned 8.0.29/9.0.18 runtime packs, which the 10.0.401
SDK no longer accepts for Photino.NET's net8/net9 targets, so it no longer
built; regenerated with fetch-deps.
Bumps the tome input to 011a69e: 10-byte device serial (Audible denies
licenses to 20-byte ones since 2026-09) and denial reasons in the error.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
mercury drops the sd-image installer profile (rescue tools, ZFS), linux-firmware,
docs, installer tools and the nixpkgs registry pin — every deploy rewrites the
changed closure onto the SD card.
deploy switch/boot/test now passes --use-substitutes on all hosts: targets pull
cache hits themselves (compressed) instead of the laptop pushing uncompressed
NARs over ssh.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds turtle-beach-swarm-re as a flake input and enables its release
module, so a warm reboot from Windows no longer leaves the KB7 showing
Swarm's last streamed frame.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
usbmon + wireshark group for USB capture, and uaccess on vendor 10f5
USB/hidraw nodes for QEMU passthrough and the RE tools.
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>
run0 waits on systemd over D-Bus, and the 25s default call timeout ran
out while the polkit dialog was still open. Scoped to an alias so a hung
systemctl does not also wait 5min.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Named profiles keep their own .env, so coder kept the old OpenCode key and
librarian, with no OPENCODE line at all, was left on an exhausted pool entry.
The OpenCode key now goes to every profile; other keys only to profiles that
already carry them.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011XkDJkdDzneVVdnkZfrvTY
Tells the console Hermes' user and AI peers, as hermes-agent.nix writes
them into honcho.json, for the "what luna sees" preview.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Honcho now exports Prometheus metrics (METRICS_ENABLED) for the
console's token, cost and latency views. The console prices each stage
with the models read from honcho-api's own environment, and reads
Honcho's OpenRouter key as a systemd credential for actual spend.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Own repo (darman/honcho-console) as a flake input; its module runs on
mars with a SELECT-only grant on Honcho's database, at
http://mars.orbit.sol:8095 with the port closed on the LAN.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A honcho.json host block without observationMode counts as a legacy config
and gets "unified" (ai observe_me off), so the seed and everything luna
says were never derived.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
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>
- prepare-dirs writes hosts/mars/luna-soul.md over her SOUL.md on every
start: a short anchor (identity, register, non-negotiables). Her own
approved edits no longer stick; a broken one corrupted the file on
2026-09-24 with GLM tool-call markup.
- luna-soul-full.md keeps the full persona. It seeded her Honcho
identity (hermes honcho identity), minus the voice rules, which moved
to luna-voiced.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- prepare-dirs pins memory.provider = honcho in Hermes's config.yaml and
merges the local server into honcho.json (workspace/aiPeer luna, user
peer darman). Only those keys are asserted; Hermes's own additions stay.
- The SDK installs itself into the image's durable lazy-install target
(/opt/data/lazy-packages), so nothing to package here.
- Hermes starts after Honcho but doesn't require it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Honcho v3.2.2 (API + deriver containers) on loopback, with native
Postgres/pgvector and Ollama embeddings. Not wired into Hermes yet;
Mnemosyne stays her memory provider until the switch.
- Peer auth over the Postgres socket: the containers run as the host's
honcho uid, so there is no DB password.
- Embeddings: qwen3-embedding:0.6b (1024 dims, fixed at first migration).
The API runs configure_embeddings on start to resize Honcho's default
vector(1536) schema.
- LLMs via OpenRouter: glm-5.3-flash:nitro at minimal reasoning for the
deriver and quick dialectic, deepseek-v4-pro for deep dialectic and
dreams (which maintain the peer cards).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Luna went silent after the OpenCode Go subscription lapsed, and a new key
still 401'd after deploying it.
- Rotate opencode_go_api_key.
- prepare-dirs writes it into Hermes's .env like the other keys, so a
stale line there can no longer override the container env.
- Changing the key restarts prepare-dirs and the container; podman only
reads --env-file at start.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CLD panel next to luna showing the 5-hour and weekly usage windows (plus
any per-model weekly caps), polled every 5 min via a new `claude-usage`
script that reuses Claude Code's OAuth login and refreshes it when expired.
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>
API_SERVER_ENABLED/HOST/PORT in the env template, bound 0.0.0.0 for the
same reason the dashboard is: the container is --network=host and nothing
opens 8642 in the firewall, so it is reachable over tailscale0 only.
API_SERVER_KEY has to go through .env like the ElevenLabs key did
(461d6cc). Confirmed the same way: with the key set only in the container
env the server came up — so API_SERVER_ENABLED/HOST *did* pass through —
but answered every bearer token with "Invalid gateway API key". The s6
passthrough list carries the plain API_SERVER_* vars and drops the secret
one. The .env write is now a write_env_key helper so both keys share it.
restartUnits gains prepare-dirs, which is what writes the .env line; the
container then has to come back up to read it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
prepare-dirs is an inactive oneshot and switch-to-configuration only
restarts active units, so the .env delivery added in 461d6cc deployed
cleanly and then did nothing — the container kept running with the old
script until something else restarted it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The image's s6 init hands the gateway a fixed set of container env vars;
ELEVENLABS_API_KEY is not one of them, so --env-file alone left
text_to_speech reporting it unset even though the var was in the
container config and the resolver read os.environ fine.
prepare-dirs now merges the single line into ${hermesHome}/.env
(HERMES_HOME=/opt/data), as uid 986 so root never writes into hermesHome,
with the key on an fd opened as root so it stays out of argv and the journal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
podman reads --env-file only at container creation, so re-rendering
hermes-agent.env left the running container on its old environment and
TTS kept reporting ELEVENLABS_API_KEY not set after the key landed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ELEVENLABS_API_KEY is the only env override ElevenLabs takes; provider,
voice_id and model_id are config.yaml-only. Seeds tts.provider through the
existing prepare-dirs yq merge so it defaults to elevenlabs without pinning
it — KittenTTS stays provisioned as the offline fallback.
Needs elevenlabs_api_key in secrets/mars.yaml before the next deploy, or
sops-nix fails activation on the missing key.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
origin/master gained PRs #7/#8 on a base that never received the three
local commits, so it still carried FORWARDED_ALLOW_IPS="*" and lacked
jupiter-reachable.service.
Conflict in hosts/mars/hermes-agent.nix resolved by keeping this side's
narrower dashboard trust (trusted_proxies + HERMES_DASHBOARD_PUBLIC_URL,
fec8d7c) and dropping the wildcard it replaced, while keeping both
KittenTTS env vars and PR #8's PYTHONPATH.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Telegram voice-note transcription on mars, no cloud: faster-whisper 1.2.1
(ctranslate2 CPU) installed into the SAME kittentts side-venv via
stt-requirements.txt (appended by the provisioning unit; input stamp
extended). Second auto-patchelf pass RPATHs the newly installed
ctranslate2/av binaries. Container env gains
PYTHONPATH=/opt/data/kittentts-venv/... so hermes's local-stt provider
imports faster-whisper in-process (the image venv is read-only for uid
986, so the built-in lazy installer cannot work there).
Validated live: Telegram voice note transcribed on CPU (i5-4460,
lang=en, text exact) with the same pins.
The venv's python is Nix python313 (also inside the container, via the
/nix/store mount), whose loader never searches /usr/lib, so PyPI manylinux
wheels failed with "libstdc++.so.6: cannot open shared object file".
auto-patchelf the staged venv against gcc-lib + zlib; RPATHs are store
paths, so the delivered copy works in the container too. The lib paths join
the input hash so a toolchain bump rebuilds.
Verified locally with mars's exact uv/python313: construct offline from the
staged hub layout and synthesize with voice "Luna".
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
uv reads the version from the wheel filename, and the fetchurl store path's
hash prefix made it reject the file ("invalid version"). Expose it via a
linkFarm under kittentts-0.8.1-py3-none-any.whl.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- pkgs.setpriv does not exist (mars failed to eval); setpriv is in util-linux.
- `sh` is not on a unit's PATH; call ${pkgs.runtimeShell} directly.
- Drop `env HOME="$hermesHome"`: unset shell var under `set -u`, and unneeded.
- Staging dir 0700 -> 0755: the uid-986 delivery step could not read it.
- Rebuild the staging venv whenever the stamp misses, else an input bump
delivered the old venv under the new stamp.
- Add diffutils: `cmp` in the fast path was missing, so every boot re-delivered.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
1: stageVenv declared with the other path vars before the early-exit
reads it (set -u no longer kills the self-heal path on later boots).
2: delivery is rm+cp as the container uid via setpriv — no mv -Tf
rename-replace on a non-empty directory (that failed every
re-provision); chown -R step gone with it.
3: network-online.target back in after= — this unit CAN download at
boot, so uv must not run pre-network.
4: root never writes into hermesHome at all. refs/main, snapshots and
venv staging happen under root-owned staging; delivery drops to uid 986
via setpriv, rm -rf + cp -a from the staging sources. Symlink-takeover
of refs/install targets and the rm/cp race on the delivered tree are
structurally gone.
Minor: sanity check now constructs KittenTTS against the staged hub
tree, not just imports; fetchurls pinned to the commit sha matching
refs/main; fast-path comment matches behavior (staging venv skips
rebuild on damaged delivery).
Review round 1 fixes:
1+2 (security): build+stamp+uv-cache moved out of hermesHome into a
0700 root-owned staging dir (/var/lib/hermes-kittentts). Root executes
python only from the staging venv after an offline import check;
finished artifacts are copied into hermesHome and the stamp compares
delivered copy vs staging byte-for-byte. No root code path follows
anything the container can write, so no stamp symlink-takeover and no
planted-wheel-in-cache execution.
3: HF snapshot dir is the real commit sha with refs/main written, so
hf_hub_download resolves the seeded cache offline (was 'kitten', never
found).
4: container env pins HF_HOME=/opt/data/kittentts-hf + HF_HUB_OFFLINE=1.
5: stamp keyed on the full input set (requirements+wheel+stub+models),
not requirements alone; deletion of a delivered file drops out of the
fast path and re-provisions cheaply.
6: orders after/requires hermes-agent-prepare-dirs.
tts.provider: kittentts, voice Luna, mini-0.8 model (80M params) per
darman. CPU-only onnxruntime inference — no GPU on mars, validated ~1x
realtime on a weaker dev box. Upstream's misaki[en]+spacy declaration is
deliberately not honored (pulls torch→CUDA, 5.6 GB); a .pth-registered
stub satisfies the dead misaki import, failing loudly if it is ever
actually used. HF model cache pre-seeded from hash-pinned store paths so
HF_HUB_OFFLINE=1 runs with zero boot-time network. Provisioner shaped by
the Mnemosyne review: uv venv --clear, root never executes python from
the venv, ownership to the container uid before anything imports, stamp
checked against venv+model files.
jupiter resolves to its tailnet IP, but network-online.target fires on
the LAN link before tailscaled routes tailscale0. Hermes'
RequiresMountsFor pulls the mount at boot, so it failed with ENETUNREACH
and took prepare-dirs, the Mnemosyne link and the container down with it,
with nothing retrying. The mount now requires a oneshot that polls
jupiter:445 for up to 5 minutes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Replaces FORWARDED_ALLOW_IPS="*" with dashboard.trusted_proxies pinned to
neptun's tailnet IP (merged into config.yaml by prepare-dirs via yq) and
pins HERMES_DASHBOARD_PUBLIC_URL so the OIDC redirect_uri stays https.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- fetchPypi: sdists are published underscore-named, so the hyphenated
pname 404'd.
- sitePackages is relative; prefix the env path so the unit's -d check
and the symlink point at the store.
- hermes_memory_provider ships in mnemosyne-memory, not mnemosyne-hermes;
move the import check accordingly.
- Drop the unused /opt/data/mnemosyne-env mount, the orphaned
requirements.txt, and trim comments.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Supersedes the venv-based fix attempt in 694317a entirely: replaces the
runtime side-venv with a python3.withPackages derivation
(pkgs/mnemosyne-env.nix: mnemosyne-memory 3.15.1 + mnemosyne-hermes 0.5.0
via fetchPypi, base deps only). Env mounted :ro into the container; the
oneshot only writes the plugins/mnemosyne symlink to the env's
site-packages passthru — a canonical store path valid on both sides. No
runtime fetch, no stamp, no root-executes-luna-writable-code, no
host/container path mismatch, no config.yaml sed.
Mnemosyne isn't bundled with the official image; third-party PyPI plugin.
Built as a side venv + plugin symlink inside hermesHome so it lands
inside HERMES_WRITE_SAFE_ROOT and survives image rebuilds. Pinned
requirements captured from a validated live install.
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>
An apostrophe in a condensed comment closed the single-quoted jq program,
so the unit script failed to build.
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