The forwarded log named the cause in one line: the prompt cache limit
defaults to 8192 MiB. llama-server saves the full KV state of every idle
slot it evicts, 112 kiB per token, so RSS climbed about 170MB per
distinct prompt until the deployed server held 7.9GB for a 1.1GB model.
Measured on homesrv today, uncapped versus `--cache-ram 512`: RSS
plateaus at 932MB from the fourth distinct prompt instead of climbing.
The task's leading guess was wrong. `-ngl 99` costs almost no RSS,
because RADV keeps device memory outside the process. Numbers and method
in docs/evals/2026-08-03-llama-prompt-cache.md.
`-c 4096` is untouched. The knob is `phraser.cache_ram_mib`, unset means
512, negative passes no flag for a llama-server too old to know it.
The deploy still runs the old image, so the box keeps its 8 GiB default
until mavend is rebuilt.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Owner's correction. It is the same host CLAUDE.md already calls workpc, and
two names for one machine read as two machines. The dated eval file keeps the
old name: a measurement is never edited after the day it was taken.
morning_routines, feeds, crawl.on_demand and netscan.enabled in
deploy/mavend.json; -ambient-token on mavweb, interpolated from a gitignored
/.env. All verified on the box: the dispatcher builds the morning plan,
/api/ambient answers 401/201, the crawler reads a named page, netscan finds 3
devices, and /events fills with scan:lan and ambient:notif.
Filed 482 (ambient reads a notification's wall clock as UTC). Corrected 479:
both capabilities work once configured, so it is not a routing defect.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NJYcaBiuny9UGSpFweQVQ1
8080 is taken several times over on this box, and the container name is the
only thing addressing it, so the port is ours to pick.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Owner's call. The code default stays off — no `search` block still means no
query leaves the LAN — but the deployed config now carries one, so a question
that is not about him reaches SearXNG before it reaches the ZIMs.
Nothing runs at http://searxng:8080 on homesrv yet. That is the designed
degradation and not a broken turn: an unreachable instance falls through to
Kiwix and she never says the search failed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The kuma service_down nudge cannot name the service. mavpoll folds the whole
monitor_status gauge into one boolean fact keyed `service_down`, and the
phraser names a service only when the fact key is the service name, so the
message is always the generic "a service on homesrv is down". Every fifteen
minutes, with nothing to act on. A fact per monitor is the real fix and it is
filed as Vikunja #444; this is what to do until then.
`disabled_rules` in mavend.json subtracts from loop.DefaultRules by name.
Config only subtracts — rules stay code, the set stays canonical and ordered
as written. A disabled rule is not gathered for either, since the gatherer
derives its key set from the rules it was given. Unknown names are ignored so
deleting a rule cannot brick a config that still lists it, and the boot log
says what was dropped, because a rule that vanishes silently looks exactly
like a rule that is broken.
Deploy turns service_down off.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
Two things that were each half-done.
The prompts handed the model copyable examples. chatSystemPrompt lost its
openers this morning and the "Я подумала, что" tic went with them, but "не
забыл ли я" appeared in its place: the removed example had been suppressing
the masculine self-reference by accident. Two predicatives are not enough
signal, so the rule is now stated as morphology (-ла) rather than as a pair of
words — a suffix rule generalises where an example only gets copied.
querySystemPrompt had the same defect and gets the same treatment; its "вот что
я нашла: " opener is deliberate and stays.
internal/kiwix had no caller. actions_query.go said "once internal/kiwix is
wired into this chain" and that never happened. It is wired now, between the
notes pass and the web source: everything of his answers first, and only what
is left over is looked up. Off unless a `kiwix` block names a server and a book.
Reading the search snippet does not work. Kiwix builds it from wherever the
keyword matched, which on Wikipedia is the navigation box at the foot of the
page — the first version of this answered "что такое фотосинтез?" by reciting
"Ecological economics Ecological footprint Ecological forecasting …". Client
grows an Article method; the head of the article is the lead paragraph, which
is the definition the snippet was meant to be. Verified on the box: the same
question now answers correctly off the ZIM.
Only the rewritten query leaves the process. A test asserts it: a turn carrying
a stored note must not put that note in the search string.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
Direct egress to api.telegram.org does not work from homesrv, so every
away-channel send timed out and the service_down nudge retried once a
minute forever. The sink already had a Proxy field wired to
http.Transport.Proxy; nothing had ever set it.
The relay is the x-ui socks inbound on the host, port 10808, addressed from
the container as the maven_default bridge gateway. That also needs a ufw
rule, because the bridge subnet is not otherwise allowed to reach a host
port and the SYN is dropped rather than refused. The rule is recorded in
the config next to the address, since the address alone is not enough to
reproduce this on another box.
Verified: five minutes after restart, zero send errors where there was
previously one per minute.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
The results channel was sized by the number of hosts while each worker
sends once per open port, so a subnet with more open ports than
addresses filled the buffer and blocked a worker forever. Nothing drains
the channel until wg.Wait returns and the sends have no ctx.Done case,
so the calling turn hung for the life of the process. Size it by probes.
Three more claims the scanner could not back. MaxHosts was spent in
order, so the second of two configured subnets got two addresses out of
254 with nothing logged. A run cut short by the cap or the deadline came
back indistinguishable from a complete one, and the shipped defaults
never fit the budget, so every scan was silently truncated at the top of
the range. Scan now reports truncation, targets are taken round-robin,
and the default rate and the budget are consistent with a /24.
The spoken reply read dotted quads out loud on the voice path. It now
says how many devices and what shape, and writes the address list as a
note, which is also the only record that Maven put packets on the LAN.
The network noun is matched whole so posetil is not a scan, the rate has
a stated ceiling, and a repeat question inside two minutes reuses the
answer.
Both query sources claimed the turn when the capability was off, which
let an unconfigured scanner and an unconfigured house swallow questions
that used to reach recall. Both now fall through.
Found in review of #81 and #80.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
internal/netscan/ discovers hosts on the network Maven is configured to look at:
a TCP-connect scan (net.DialTimeout, no raw sockets, no privileges) plus a read
of the kernel's ARP cache. Wired as a read-only query source, "network", so
"какие устройства в сети?" is answered by a scan instead of by whatever old note
happens to be nearest.
Scanning is a read, but an unbounded scanner on a home LAN is noisy and easy to
point somewhere it should not go, so the package is built around four bounds:
- Scan takes NO target argument. The range comes from the config block and
from nowhere else, so there is no exported way to scan an arbitrary prefix
and nothing an utterance, the router, or a scanned host says can retarget
it. That is asserted directly: the test watches every address handed to the
dialer and fails if one falls outside the configured prefix. The ARP cache —
the one input the network itself populates — is filtered to the configured
range for the same reason.
- Every configured CIDR must be private (RFC1918 / CGNAT / link-local) and no
larger than 1024 addresses. 8.8.8.0/24, 0.0.0.0/0 and 10.0.0.0/8 are refused
at config load, not after the packets have left.
- Rate-limited to a configured connections-per-second across the whole scan,
so it looks like background traffic rather than a portscan.
- Bounded in total by MaxHosts, a per-connection timeout, a 20s turn budget
and the context; a canceled scan stops dialing immediately.
Off unless configured: dark without "enabled": true, and applyDefaults
normalises a disabled block to nil. deploy/mavend.json carries it disabled.
BLUETOOTH IS NOT SHIPPED, AND IS BLOCKED, NOT SKIPPED. The plan's other half
(internal/bluetooth/, RSSI presence probes) needs a bluez stack that is not
here: bluetoothctl and hcitool are not installed, bluetoothd is not installed,
the bluetooth unit is inactive, and org.bluez is not on the system bus. hci0
exists as a kernel device and nothing can talk to it. The docker deploy is
further away still — it would need host networking, the D-Bus system socket
passed in, and CAP_NET_ADMIN. Writing an exec wrapper around a binary that does
not exist, against an output format nothing here can produce, would be a guess
dressed as a feature. It needs a decision about privileging the container before
any of it is worth writing.
Vikunja #257
A `smarthome` block points Maven at a Home Assistant instance. She reads its
entity states to answer "что включено дома?", and every controllable device
becomes a PROPOSED row in the existing act allowlist — cmd
["smarthome",<entity_id>,<service>], scope smarthome:<domain> — so nothing new
had to be invented for the mutating half. ProposeTool/EnableTool/DisableTool,
tool.Matcher and the confirm turn are untouched; one branch in Executor.Exec
routes such a row to the client instead of exec, and "smarthome" is never run as
a binary. This is the same trick overnight/mcp-tools used for #251, on purpose.
Discovery only ever PROPOSES, and every control row is destructive=true: there
is no read-only way to turn the heating off, so flipping something in his flat
always costs a confirm turn and always had to be enabled by hand on /tools,
behind step-up.
The entity and the service come from the row he enabled, never from the
utterance — Exec drops the spoken tail for a house row. A router that misheard
can pick the wrong lamp; it cannot compose a target of its own. The service is
checked against the domain's table on the way out too, so a hand-edited cmd
column cannot reach an arbitrary Home Assistant service. set_brightness and
set_temperature are deliberately absent: a spoken number the router got wrong is
a wrong act on real hardware, and on/off is the whole of what a voice turn can
defend.
The read side is a query source ("home", before calendar and the recall passes)
so "что нового дома?" is not answered from an old note. Its matcher needs a
house marker plus an ask plus a device word and bails out on weather wording,
because "какая температура на улице?" belongs to the weather source.
Off unless configured: the block is dark without "enabled": true, and
applyDefaults normalises a disabled block to nil so "off" stays in one place.
deploy/mavend.json carries it disabled, with the token as ${HA_TOKEN}.
NOT shipped, and not faked: MQTT / Zigbee2MQTT (plan steps 2 and 5) and the
sensor-to-fact and presence-probe pipelines. There is no broker and no Home
Assistant anywhere on this network — 8123 and 1883 are closed on every host in
192.168.1.0/24 — the module tree is vendored so a paho dependency cannot be
added offline, and Home Assistant already fronts Zigbee2MQTT where it exists.
Writing a sensor pipeline with no sensor to test it against would be a guess.
Vikunja #256
Second half of the MCP client: the tools the manager discovers become rows in
the existing act allowlist instead of a parallel capability system.
An MCP tool is encoded in the columns that already exist — cmd
["mcp",<server>,<tool>], scope mcp:<server> — so no migration, and
ProposeTool/EnableTool/DisableTool, tool.Matcher and the confirm turn need no
changes. One branch in Executor.Exec routes such a row to the manager instead
of exec, and "mcp" is never run as a binary.
Discovery only ever PROPOSES. destructive comes from the inverse of the MCP
readOnlyHint, so a tool that does not promise to be read-only inherits the
confirm turn, and enabling stays on /tools behind step-up.
Voice args are positional and MCP args are named, so CallPositional binds only
what it can defend: no required properties runs bare, and a read-only tool with
exactly one required string or number gets the tail. Everything else refuses
with ErrNeedsArgs rather than guessing. The read-only condition was learned
against the live Vikunja server: update_task requires only task_id and takes
the rest as optional, so one guessed argument blanked the fields it did not
mention. A partially-filled write destroys what it omits, so a mutating tool
never receives a guessed argument.
Also: a read-only mcp_servers IPC method and an "MCP servers" card on /tools
showing transport, target and state, with the trust level of a local target
spelled out. There is deliberately no call-a-tool IPC method and no run button,
so mutation keeps exactly one path.
Vikunja #251
The digestion tick already runs the pattern detector over all recorded
events (67563ed) and writes a proposed_routines row. What was missing is
the other half of #247: a proposal that nobody is at the mic for reaches
nothing but the /routines page, so a pattern noticed at 03:00 is only
seen if he goes looking.
This wires the tick's proposals into the existing care-delivery path
rather than a second channel: sev1 nudge, loop.Gate, dispatcher, same
routing table as an accepted routine. Restraints, since a feature that
speaks unprompted is the easiest way to turn Maven into a nag:
- off unless configured — the new pattern_proposals block, absent by
default, and deploy/mavend.json ships notify: false;
- at most one announcement per tick however many patterns surfaced;
- at most one per cooldown (24h default) across all pairs;
- sev1, so quiet hours, away and snooze suppress it;
- suppressed means dropped, not queued — /routines still has it;
- once per pair for good, since proposed_routines is
UNIQUE(action, object) and the row survives dismissal.
The body is pattern.PhraseRoutine's literal Russian, not LLM-worded, so
an inferred routine cannot arrive describing something never observed.
Also raises pattern.MinEvents from 3 to 4 — the interval-quality item on
#43. Two intervals with a ±50% band is a coincidence with a mean, not a
pattern, and now that a scan of all history can announce itself the cost
of a false positive is a permanent dismissal of that pair.
Stock Qwen3-1.7B, not the CPT'd one — that training is still running. It won
on both fixtures we have, measured tonight on an otherwise idle box:
routing, 77 RU cases, intent-only: 67.5% vs 59.7% for Qwen3.5-0.8B
talk fixture, 27 cases: 20/27 vs 11-17/27
It also beat Qwen3.5-2B, which is 20% larger, on every routing column.
Two other things came with it:
n_ctx goes 2048 -> 4096. This is a Thinking variant, so reasoning tokens need
the room, and 4096 is the context every score above was measured at. Shipping
2048 would ship something nobody measured.
The doc now says not to bother with sub-500M models, because I checked and they
are not close. LFM2.5-350M routes at 5.2% — worse than guessing among 7 intents
— and answers "столица Франции?" with "Сторзит", which is not a word. The 230M
replies to Russian in Spanish. Their published IFEval and BFCL numbers are good
and they are all English.
Note the routing gain needs the LLM router actually wired on to show up. It is
still nil, so this commit buys the phrasing improvement today and the routing
improvement when that lands.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
DEPRECATION, flagged not asked: LLM-phrased nudges are no longer the default.
LLMPhraser.PhraseNudge now returns a hand-written Russian template. The model
still phrases chat, queries and reminders — only nudges moved.
Why: measured over many runs, Qwen3.5-0.8B wrote formal "вы" and plural
imperatives, used masculine self-reference, and invented facts and units
(90-95 seconds to boil an egg). A nudge is five words of known content, so
generation buys nothing and risks the persona every time. Templates score
15/15 on the nudge fixture, the model 11-13/15.
Nothing is deleted: the prompt, the fallbacks and the whole LLM nudge path
stay. Set phraser.llm_nudges = true in deploy/mavend.json to get them back.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The e5 embedder puts every cosine in one narrow band (0.79-0.89), so the
absolute query_min_score gate cannot tell a real hit from a made-up
question: any value under the band answers everything, any value above it
answers nothing. False recall was 5/5.
New gate asks whether one note is clearly the best instead: top1 - top2 >
delta. New query_min_margin config knob, default 0.008, read off the sweep
in the recall harness. The absolute floor stays as a second check.
On the recall fixture with e5: answered 72% -> 68%, false recall 5/5 -> 1/5.
Vikunja #359
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The two things that made this unsafe are fixed: the router can now
refuse, and slot extraction runs on its decisions.
On the held-out fixture it gets 63.2% of intents right against the
classifier's 50.0%, with no route errors. It costs about a second a
turn instead of 30ms.
The flag is a pointer now, so leaving it out of the config means on
and only writing false turns it off.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
The old model was a symmetric paraphrase model, so it scored "do these
look alike" instead of "does this note answer this question". Also fixes
the file mismatch: the Makefile, the deploy config and both evals now all
name the same quantized file, and the quantized one is what gets measured.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Wires cmd/mavend/voice.go to build the LLM router when the operator asks
for it. Default false, so nothing changes on the deploy box.
Look at pickLLMRouter: the flag on with no llama-server logs one line and
keeps the classifier, it never fails a turn.
The default stays off until the router can refuse (#359) and the extractor
runs on LLM decisions — both noted as TODOs in config.go.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
Turns on the existing digest path on homesrv: a 30m batching window, at most 5
items per flush, and a severity ceiling of 2 so anything more urgent still
goes out immediately instead of waiting for the batch.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X5JApcrCRVGmqrxnhynSik
The most load-bearing decision in the project was stated four incompatible
ways: the docs said Qwen3-1.7B, deploy/mavend.json said Qwen3.5-2B, the repo's
models/llm/ held an LFM2.5-1.2B gguf, and five code comments still said LFM.
Answering "which model is deployed" meant re-deriving it from scratch every
time.
Two facts the review missed, found while resolving it:
- /mnt/hdd1/llms is bind-mounted over /opt/maven/models/llm, which shadows the
repo's models/llm/. The LFM2.5 gguf sitting there was never loaded by
anything, so it was not evidence of the deployed model at all.
- That library holds Qwen3.5-0.8B, -2B and -4B, and no Qwen3-1.7B. The config
pointed at a file that does exist; the docs' Qwen3-1.7B was the stale claim,
the reverse of the assumed direction. Qwen3-1.7B is the CPT target, and that
training is still in flight (Vikunja #122), so no such gguf exists yet.
phraser.model_path moves to Qwen3.5-0.8B (Q4_K_M) — the smallest checkpoint on
disk, chosen for latency, and relevant to whether the LLM router is affordable
on this box. Docs and comments now say the same thing in one voice: 0.8B
resident now, CPT'd Qwen3-1.7B as the target, and the bind-mount shadowing
written down so the next reader does not mistake models/llm/ for ground truth.
Comments name the model, never a filename, so a swap stays a one-line config
change.
n_gpu_layers: 99 is correct and stays — compose passes /dev/dri and the render
gid for Vulkan offload to the Vega iGPU. CLAUDE.md's "CPU-only" was the stale
half of that contradiction and is corrected.
phraser.go also dropped a wrong "sub-1b, prompted not trained" size claim: the
target is trained end-to-end (RU CPT + joint persona/router SFT).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01X5JApcrCRVGmqrxnhynSik
WriteFactReq gains an optional Subject field (empty = old behavior,
no CoreAPI signature change) and the IntentFact handler now passes the
fact's key as its resolution subject, so voice-tapped facts flow into
the Vikunja #279 enrichment queue automatically.
Also: deploy/mavend.json's phraser was pointed at a 4B model with
n_gpu_layers=99, which OOM'd under memory pressure and left a zombie
llama-server child. Swapped to the 2B Qwen model matching the intended
resident-model size, keeping GPU offload.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_018ghELqYhZNLub2TXGMazqA
Add a read-only /ecosystem page that consumes the sibling services'
JSON APIs (Nexus entities, Praxis attention, Hexis capabilities),
fetched concurrently with honest per-panel error states. Siblings stay
headless — mavweb is their human surface (arch §16). Wired via mavweb
-nexus/-praxis/-hexis flags; mavweb joins the ecosystem compose network.
Fix mobile horizontal overflow across all pages: .content is a flex
child with default min-width:auto, so it refused to shrink below the
tables' intrinsic width. min-width:0 lets wide tables pan inside .scroll
instead of dragging the page sideways. Verified via CDP geometry check
(scrollWidth === clientWidth at 430px).
Also includes in-progress Ethos UI redesign, ecosystem deploy compose,
and planning docs.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Daemon side of Decision B: parse {"response","mood"} across the 4 consumers
(replier, nudges, reminders, chat), fall back to legacy formats. Drop the
LLM router — the classifier handles routing; replier/phraser share one
llm.Client (timeout 20s->60s). llm.Client reads reasoning_content when
content is empty (thinking models).
Docs: TTS piper-student plan (OmniVoice teacher -> piper student, from
scratch, phoneme-first). CLAUDE.md training guide.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
"voice unavailable" in the web UI: mavweb dials mavend:9100, but the container
mavend.json had no voice block, so mavend never bound 9100 (worked pre-docker
because the host's ~/.config/maven/mavend.json had one). Ported that block:
enabled, bind 0.0.0.0:9100 (not 127.0.0.1 — mavweb is a separate container),
lang ru, stt/tts worker sockets, onnx embedder.
Enabling the embedder surfaced a second bug: the router needs onnxruntime 1.26,
but deps/lib only carries dangling symlinks to it (absolute host paths, not in
the image), so the only libonnxruntime present was piper's 1.14 (copied in) →
"ORT API base: 2", crash loop. Fixed the Dockerfile to ship the real 1.26 .so
and stop copying piper's .so into the shared lib dir (piper finds its own 1.14
via $ORIGIN + exact soname, so TTS is unaffected).
Verified: mavend "onnx embedder loaded (384 dim)", "voice listening on :9100",
stack stable.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Compose stack replacing start-maven.sh's bare `&`-backgrounded processes.
Single multi-stage image builds all six daemons (CGO + prebuilt native libs
from deps/); compose runs one container each with a different command. Only
mavend mounts the encryption key (env_file, gitignored) and the db volume; the
modules mount just the shared unix-socket dir and read-only models — so the
"key-free modules" boundary is OS-enforced (separate namespaces), not just a
code convention. IPC stays unix-domain over a shared volume: zero code change,
paths move to /run/maven. Encrypted db at rest on a named volume, decrypted
working copy in tmpfs (RAM) per the at-rest encryption landed earlier.
Validated: `docker compose config` clean, mavend.json parses, all daemon flags
confirmed. NOT build-tested (no docker/GPU in authoring env) — deploy/README.md
lists the host-dependent tweak points (GPU passthrough, onnxruntime path,
cross-container voice bind, netdata host).
Chosen Docker over interim systemd units per the "dockerize soon" call — no
throwaway supervisor built.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
claude
kami