The replier and the meeting summariser were the two call sites without a
GBNF. Both are exactly the shape that makes a Thinking variant answer with
its reasoning as prose, and neither had anything downstream that could
remove it.
The replier already parses {"response","mood"}, so it now sends the phraser's
grammar for that contract, exported once as phraser.ResponseGrammar so the
two definitions cannot drift.
The summariser stays text-in/text-out. The JSON wrapper is attached and
unwrapped in the daemon's Completer, so internal/capture is unchanged and a
Completer without a grammar still works.
The simulator told routing from phrasing by "has a grammar", which stopped
being true here; it now looks for the intent enum.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The offline encyclopedia was the only world source, and it reads what was true
when the ZIM was built. A self-hosted SearXNG now asks first and Kiwix is the
fallback for an empty result, an unreachable instance or no line out. Owner's
ruling, 2026-08-02.
internal/websearch is deliberately thin: no rewriter (SearXNG ranks through
real engines, so the Russian question goes out as he asked it), no page fetch,
no cache. It cannot read the store, so only the query string can leave the box.
The personal boundary is unchanged and still sits above this source, so a
question about him is never searched.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BFaeSbLMEVG5ey8tejU3y2
The budget was cut to 60s because a five-minute evaluation held the single
llama-server slot, and a voice turn arriving mid-evaluation waited behind it.
That collision is now solved where it belongs: the background client yields the
slot while a turn is in flight.
With the gate in place the short budget only truncates a Thinking model
mid-synthesis, which costs an observation and saves no latency on any real turn.
Kami's call, 2026-08-02.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BFaeSbLMEVG5ey8tejU3y2
Three documentation changes and one deletion.
CLAUDE.md and AGENTS.md gain the Nexus/Praxis/Hexis sections that were written
last session and never committed: what each service owns, where Maven's client
for it lives, and the rules that are not negotiable.
The p50 latency figure was wrong in two files. CLAUDE.md said the cascade costs
2.7s and that the LLM router is 90x slower than the classifier. Both come from
the bakeoff table, where the number is contention on a shared llama-server, not
the model. ROUTING-EVAL-31-07-2026.md line 61 says so and measures the router at
p50 825ms / p95 1.2s / max 3.0s. Corrected in CLAUDE.md, and the bakeoff table
now carries a header pointing at the routing eval for absolute latency. Latency
work was about to be planned off a number that was never real.
HANDOFF.md is deleted. It described work sitting on fix/integrated waiting for a
fast-forward onto overnight/eco-versioned-traces. Neither is true: master
contains that tip plus 22 commits, and both branch pointers are stale. The three
live defects it recorded move to PLAN-DETERMINISM-02-08-2026.md, which is now
the only planning document.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01BFaeSbLMEVG5ey8tejU3y2
"во сколько у меня встреча" fell past the calendar, reached Kiwix, matched
an article on the 2015 CPISRA World Games and came back phrased as his
meeting. Inventing is worse than refusing, and as `system` this used to say
"пока не умею".
A new source sits between the notes pass and Kiwix: a question carrying a
possession marker ("у меня", "мой", "my", "do i have", "did i") that his own
data did not answer has no answer outside it either, so the walk stops there.
The markers are possession, not first person, so "как мне сварить борщ" still
reaches the encyclopedia.
This is also where CLAUDE.md's privacy line lands: only the utterance may
leave the box, and a question about him carries his life in the utterance.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
wireVoice runs before the tick loop exists, so it could only be handed
the bare store adapter — and that adapter answers DayPlan with "not
available via direct store API", because a day plan is assembled by the
tick loop and is not a table to read. So queryDayPlan, which the query
chain reaches for "какие у меня планы на сегодня", failed for every
caller on the deployed daemon.
main already back-patches the other direction (daemonAPI.chatFn =
handler.handleText). This is the same seam in reverse, at both wiring
sites. No recursion risk: nothing in the voice path calls api.Chat.
With the plan reachable, it recited its reminders as literal JSON. The
payload unwrapper existed but was private to the phraser, so the day
plan had its own non-unwrapping copy. One owner now, store.ReminderText,
with the phraser delegating to it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
rememberTurn runs after followUpMerge, which has already inherited a
Text slot from the previous same-intent turn, so the fill-if-empty rule
pinned the first topic of a run of query turns and never released it.
"во сколько у меня встреча", then "какие у меня планы", then "а завтра?"
continued the meeting — two turns stale.
Overwrite for system and query, where Text is a topic and not a payload.
A continuation is the exception and keeps what it inherited: its own
utterance is the ellipsis, and the topic it carries is the real one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
"что у меня сегодня" and "что у меня в календаре сегодня" both routed
IntentSystem on the deployed daemon, and replySystem has no agenda arm,
so both answered "пока не умею". The calendar source that can answer
them lives in the query chain and was never reached. The fixture has
said query since ru-query-019 was written; the daemon disagreed with the
fixture and the daemon was wrong.
AgendaQueryGrammars routes them at stage 0, after the clock rules so
"какой сегодня день" keeps reaching replySystem. Intent only — which
source claims the turn stays the query chain's decision.
This is what made the follow-up continuation look like it only worked
for "what day is it". It did: the query half inherited an intent whose
handler could not answer, so both halves came back "пока не умею".
Measured on the 77-case RU fixture: full accuracy 70.1% → 72.7%,
intent-only 75.3% → 77.9%, calendar 0/2 → 2/2, clarify counts unchanged.
The eval harness wires the new grammars too, or the fixture would stop
being a measurement of the daemon.
Go's \b is ASCII-only and never fires after a Cyrillic letter, which the
first version of the pattern learned the hard way.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
The follow-up continuation re-aimed Slots.Time, but every query source
matches on dec.Utterance and nothing in the chain reads it. So the
inheritance bought almost nothing: "какие планы на сегодня" then "а
завтра?" missed day-plan's matcher and fell to the calendar, which had
parsed the day out of the raw utterance anyway.
Worse than nothing in one place. queryFactByKey runs first and claims on
HasKey plus HasTime, both of which the continuation sets, so a keyed
query continued with "а вчера?" answered "я записала это <the fact's own
timestamp>" and dropped the day entirely.
Widening the topic for every source would have made it worse still, not
better: CalendarEvents is the only CoreAPI call that takes a date, so
ten date-blind sources would have answered a question about tomorrow
with today's data. Gate instead of widen — a continuation is only
offered to a source that reads the day, and when none claims it she says
so rather than "не знаю", which reads as "nothing tomorrow" when she
never looked.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
Deployed check, second round: "привет" after "какой сегодня день" answered
with the date. followUpMerge fills an empty Text from the previous
same-intent turn, so the topic-widening added a minute earlier was reading
an inherited topic on turns that had nothing to do with it.
Decision gains Continued, set only by continuation.go and never by the
router. replySystem widens on that and nothing else, so an inherited Text
is back to being invisible.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
Deployed check: "какой сегодня день" then "а завтра?" answered "пока не
умею". The intent was inherited correctly, but replySystem keyword-matches
the utterance, and "а завтра?" contains no topic word — that is the whole
nature of an ellipsis.
So the topic travels with the session. rememberTurn keeps the raw utterance
in Slots.Text for system and query turns that have no Text slot of their
own (a stage-0 grammar fills none), and replySystem matches keywords
against utterance + Slots.Text. Dates keep parsing from the utterance
alone, which is the part the ellipsis actually restates.
Only system and query: everywhere else Text is a payload and must stay
exactly what the router put in it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
Live check on the deployed daemon: "напомни сегодня о событиях" then "а
завтра?" fires tomorrow with the text still reading "сегодня". Re-aiming
Time is not enough when the day word is also inside the payload, and
rewriting the payload needs the date's span in the string, which
ParseCalendarDate does not report.
Reminder comes out of continuableIntents until that exists. query and
system are unaffected: their Time slot IS the whole question.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
An elliptical follow-up carries no intent of its own. followUpMerge cannot
help — it inherits slots once the intent is known, and here the intent is
the missing part. So "а завтра?" went to the router, which on a 1.7B is
close to a coin flip, and the guess cost ~2.7s.
continuationDecision runs before the router and rebuilds the turn from the
previous one: same intent, same key, new day. Deterministic and free.
Three guards, all narrow on purpose. A parseable date is required, which is
what separates an ellipsis from an ordinary short utterance. Four tokens
max. And only query, system and reminder may be inherited: fact and note
would write something he did not say, and act would let a two-word
utterance re-run an allowlisted fn, which is a way to fire a destructive
command nobody typed.
A continuation is still remembered, so "а завтра?" then "а послезавтра?"
chains.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
The snooze wire had one end. "готово" and "выпил воды" both left the nudge
pending, so the auto-tuner only ever learned from deferrals and from
silence — never from the rule working.
Two entry points, because the two utterances are different acts. A bare
"готово" carries no content and is intercepted before the router, sharing
pendingNudge and the twenty-minute window with the snooze. "выпил воды" IS
content: it routes normally, writes its fact, and only then closes the
nudge (ackFromFact, after applyAction). Folding the second into a pre-route
intercept would have thrown away the thing he actually said.
ackFromFact is silent. The fact reply stands; "отлично, отметила" on top
would be her congratulating him for obeying, which is the nag she is not.
Which fact answers which rule comes from the rule's own InertWhenNoData, so
a rule added later is covered the day it lands. "да" and "ок" stay out of
the ack vocabulary: the clarify gate upstream has the stronger claim on
them, and a stray "да" must not rewrite the feedback signal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
A nudge could only be deferred from Telegram or the web UI. The voice path
had no route to store.ResolveNudge at all, so the channel she nudges on
hardest was the one he could not answer out loud.
resolveSnooze runs pre-route, right after the quiet toggle, and writes the
same `snoozed` outcome the buttons write — which also drops the row out of
RepeatUnacked, so a deferred sev4 stops re-sending every five minutes.
The window is what makes this safe to run before the router. "потом" is an
ordinary word; it only counts as a deferral when a pending nudge was sent
in the last twenty minutes, and otherwise the turn routes normally. Single
word patterns still match single-word utterances only, so "потом схожу за
водой" reports a plan instead of silencing the rule that prompted it.
Channel is not filtered: a nudge that went to Telegram is still what he is
answering when he says "потом" at the microphone.
QA-PLAN gains the two new manual checks and drops the 319 warning.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
The pre-route toggle knew "тихий режим" and every negation of it, but not
the two phrasings that get spoken most: "включи режим тишины" (the setting
named as a noun) and "сделай потише". Both fell through to the router,
which has no quiet intent, so the command did nothing at all.
Adds those as stem pairs, plus a quietWordStems list so the
negated-but-unmatched fallback recognises "хватит тишины" the way it
already recognised "хватит тихого режима".
Locks the English phrasings from the routing fixture in the test table
("turn quiet mode back on", "turn off quiet mode", "stop quiet mode") —
all three already behaved, none were covered.
"на улице стало потише" and "в тишине лучше думается" stay inert: a
single-word pattern still only matches a single-word utterance.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
The clarify gate thinned any one-word utterance to 0.3 confidence, which
trips the stage-3 gate and comes back as "не совсем поняла". That is an
English intuition. Russian packs subject, tense and gender into one word,
so "поужинал" is a complete report and "привет" a complete greeting, and
both got clarified.
thinSingleToken keeps the rule for bare nominals, where it is real ("вода"
is a fact-or-query coin flip), and spares two classes: a closed lexicon of
social and control singles, and any token carrying a verb ending. Both
tests are offline.
Fixture: false clarifies 3 → 2, intent-only 74.0% → 75.3%, full accuracy
unchanged at 70.1%, missed clarify still 1.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
`disabled_rules` stopped the loop from creating new nudges and did nothing
about the ones already sent. The sev4 repeat path does not consult the rule
set at all: RepeatUnacked re-sends any telegram nudge still at outcome=pending
every repeat_interval (5m by default), driven by store.UnackedTelegramRules.
So service_down kept arriving on a five-minute cadence after being switched
off, from a row written hours earlier — two messages after the deploy, which
is how it was found.
That cadence, not the unsealed database, is what "she keeps spamming me"
always was. The seal bug erased the acks that would have stopped it.
Filter the repeat keys against the wired rule set. Filtering on wired rather
than on the disabled list also silences a rule deleted from the code: nothing
can ack what the UI no longer lists.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
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
mavend seals its encrypted database in `defer st.Close()` when run() returns.
It had not returned since 2026-07-21. Every restart since then decrypted the
same eleven-day-old ciphertext and rolled back everything written in between:
the Telegram nudge that kept firing was a fact being un-written on each boot.
The goroutine dump named it. main → srv.Close() → ipc.(*Server).Close →
wg.Wait(), waiting on per-connection goroutines parked in readFrame. Close
shut the listener and nothing else, so the idle persistent sockets held by
mavweb, mavpoll, mavcaldav and mavmaild blocked shutdown forever. `docker
compose stop -t 60` spent the whole sixty seconds and then took a SIGKILL.
So: track the accepted conns and close them, in ipc and in voice, which had
the identical defect. Bound all three waits — the two per-server ones and the
worker wait in main — because the seal matters more than any single in-flight
call. A dropped RPC costs one reply; a missed seal costs a session.
The regression test leaves a client connected and idle, which is the case the
old tests avoided by closing the client first.
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
The grammar examples in chatSystemPrompt were full clauses — ("я подумала",
"я рада") for her, ("ты сказал", "ты забыл") for him. A 1.7B copies those
instead of generalising from them.
Observed on homesrv 2026-08-01, in one session: all three chat replies
opened with "Я подумала, что ...", and one ended "...немного тревожусь.
ты сказал" — the second example pasted onto a finished sentence, which
reads as a truncation and is not one.
Contrastive pairs replace the openers, so the rule reads as a correction
rather than a template. The him-examples are dropped; the "ты" instruction
carries that on its own, and those two produced the worst output. A closing
line tells her not to echo the instructions, because a small model treats a
quoted string as licence to reuse it.
Verified after rebuild: three chat turns, no "Я подумала" opener, no
dangling example, feminine forms intact.
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
net/http wraps every transport failure in *url.Error, whose Error() prints
the request URL. Telegram accepts the bot token nowhere but the URL path, so
a send failure wrote the live token into the daemon log. On 2026-08-01
homesrv could not reach api.telegram.org and did that once a minute for as
long as the network stayed down. The token lives in deploy/telegram.env to
stay out of the repo; putting it in `docker compose logs` undoes that.
Both error sites now go through redact. The structural branch rewrites
url.Error.URL and keeps the type, so errors.As still matches; anything else
falls back to scrubbing the rendered message. No minimum-token-length guard:
a one-character token would shred the message, but that beats leaking it.
DESIGN.md still said the classifier cascade was the path that runs today
with llmrouter wired nil, and gave the resident checkpoint as Qwen3.5-0.8B.
Both stopped being true on 2026-07-31.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
7ab9b48 committed models/stt and models/tts as symlinks to their own
absolute paths:
models/stt -> /home/kami/apps/Maven/models/stt
They came from an agent worktree under scratchpad/wt, where a link back
to the main checkout resolves. In the main checkout it points at itself.
Both paths are gitignored, so checking out that commit overwrites the
real model directories without warning and git says nothing. On homesrv
it destroyed models/stt/ggml-small.bin and the piper voice, and mavsttd
crash-looped on the missing whisper model.
The directories are host state fetched separately, per deploy/README.md.
Nothing under models/stt or models/tts belongs in git.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
One linear chain of 35 PRs, reviewed and fixed. The eleven fix branches were merged onto fix/integrated and fast-forwarded onto this tip, so the review findings land as commits here rather than on the individual PRs.
make test and make build pass.
The vendored copy of github.com/kami/hexis predated Client.WithToken:
no token field, no setter, no header hook, and an unexported httpClient,
so there was no way to attach auth from outside the package. wireEcosystem
handled that by refusing to wire Hexis at all when a token was configured,
which was the honest reading of the code but left the deployment silently
without its executing service.
go.mod already replaces the module with /home/kami/apps/hexis, and that
source has had WithToken and the Bearer header for a while. Only the
checked-in vendor/ copy was stale. Refreshed it (client.go plus the new
capability.go) and wired Hexis like Nexus and Praxis.
Two tests cover the outcome the refusal was standing in for: a configured
token reaches the wire as Authorization, and no token still wires unauthed,
because Hexis without auth is a valid deployment on a trusted box.
Also corrected the discoverCapabilities comment. It claimed the client
stamped the correlation header on Execute only; do() stamps it on every
request, and did before the re-vendor too.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
IngestMail was AuthRead and SetTaskStatus was AuthWrite, and they answer
the same question: may this module change what is on his lists? The old
argument for AuthRead — ingestion is additive, it can only produce
candidate tasks — is still true and is the weaker half, because a
compromised mail reader that can fill the review page indefinitely is
not a read.
Nothing loses access. AuthWrite outside WriteFact only requires
enrollment, which mavmaild already has, and the method does not exist
unless the operator wired a mail block.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
Two fix branches independently added a Gate to internal/llm. One is priority
between a voice turn and a background job, the other is admission control while
the resident model is swapped. They are orthogonal and both are needed, so the
swap one is now SwapGate, with SetSwapGate to install it.
Complete takes the priority gate first and the drain second. A background
request can wait a long time on priority, and counting it as in flight against
the drain that whole time would stall a swap on a request that has not started.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
capture_stop held the IPC request open for the whole map reduce, up to
twenty minutes. A voice turn that says "хватит" waited for forty model
calls before Maven said anything. Stop now returns the transcript and the
summary runs on a goroutine in the daemon's WaitGroup, on the daemon
context so a client that hung up does not cancel the only readable record
of the meeting.
With no summary and save_transcript false, writeNotes wrote nothing at
all: an hour of meeting left a blob that prunes in seven days and no
trace in the note store. The transcript is written instead when the
summary is missing. That flag decides whether the verbatim record is kept
in addition to a summary, not whether the meeting is remembered.
The wire carries the session token now, and the contract comments say
what the code does: the summary is usually absent from the stop
response, and re running a stored blob is a manual job because no method
takes a blob id. The save_transcript comment says the cost is recall
corpus rather than disk.
Found in review of #73.
Four invariants the comments claimed and the code did not hold.
The recording lived in mavend's heap as one growing []byte, doubled at
Stop when the WAV was built. Frames now go to a spool file and the
transcript is read back off disk one window at a time, so memory is flat
whatever the length.
A store failure returned before transcription ran, so a meeting over the
blob cap produced no transcript, no summary and no note. It now records
the failure and keeps going, and the spool file survives until the words
have been read off it.
The session had no owner. Any module on the write rung could call stop on
a recording it did not start and receive the verbatim words of everyone
in the room. Start hands back a token and append, stop and abort require
it.
The duration cap was only checked when a frame arrived, so a phone whose
tab was closed left the slot occupied and every later start answered
ErrBusy with a meeting from last week. The wall clock is checked in
start, status, append and stop.
Smaller things in the same pass. Append compares the frame format against
the session format, so a client that switches sample rate mid meeting no
longer has its frames concatenated under a header that lies. One failed
STT window leaves a marker instead of discarding the other twenty four.
Summarize is separate from Stop and assigns the salvaged per chunk text
before it reports the error.
Found in review of #73.
Put takes a []byte, so storing a recording meant the whole recording in
memory. A two hour meeting at 16 kHz mono is about 230 MB of WAV, and
building it from PCM held a second copy of the same size in the process
that also owns the database and the resident model. PutFile stats the
file, hashes it in a stream and renames it into place, so the peak is one
buffer regardless of length. SpoolFile hands out the scratch file it
moves from, under the media dir so it shares the same disk and the same
permissions.
Audio also gets its own per blob cap of 512 MiB. The image cap of 64 MiB
is 35 minutes of audio, which contradicted the two hour session cap: the
long meeting was exactly the one that failed to store.
audio.WAVHeader is split out of WAVFromPCM because a spooled capture
writes a placeholder header first and stamps the real length at the end.
Found in review of #73.
The docker-shaped update block in TestUpdateBlockValidatedAtStartup has
source_dir equal to install_dir and no source_rollback, which is exactly the
deployment the new validation refuses. The fixture is meant to be the good
case, so it now says how the source is rolled back.
Found in review of #69.
The on-demand crawler was built over allow_hosts plus every watched host.
webfetch reads a non-empty allow list as these and nothing else, so a config
with one watch and no allow_hosts at all silently narrowed on-demand reading
to the watched site. Every other url he pasted came back as a flat refusal
with nothing in the log to explain it. The two crawlers now take two host
lists from one crawlHosts helper.
Crawl-delay was parsed into Rules and never read. The only pacing was the
fetcher's flat one request per host per second, which cannot express what a
site asked for, and deploy/README claimed the field was honoured. Page now
waits it out between the robots fetch and the page fetch, and a delay longer
than the turn fails the read instead of hanging it.
A robots.txt that failed was treated as no rules, so a site whose server was
having a bad minute became a site with no restrictions. A 5xx now refuses the
crawl. A 404 still means unrestricted, which is what the standard says.
The refusal check matched substrings of webfetch's message text from a package
that cannot import webfetch, so a reworded error would have silently turned
into a robots verdict. internal/crawl now exports ErrFetchRefused and
ErrFetchStatus and the adapter in cmd/mavend maps the webfetch sentinels onto
them. Robots group selection picks the longest matching agent prefix instead
of the first one in file order.
queryWeb passed a claim it could not serve when no crawler was configured, so
an unconfigured deployment answered a web question with an apology instead of
falling through to the model.
Found in review of #67.
Nothing at read time told a feed item or a crawled page apart from his own
notes. QueryNotes ranked every note by cosine and the notes answer handed the
nearest five to the phraser, so "что я говорил про переезд" could be answered
out of a stranger's web page, prefixed with "вот что я нашла: ". Recall now
excludes the read sources, rss: and crawl:, and the list is one place.
The feed answer needed a different read as a result, and it needed one anyway:
it scanned the last 200 notes of any source, so a busy day of voice notes pushed
the newest headline out of the window and she said "в лентах пока ничего
нового" while the poller was working fine. RecentNotesFromSource asks for feed
notes by source, so the window holds 200 of them.
Found in review of #66 and #67.
expect_not_called could pass on a step that made the forbidden call. callPaths
concatenates per server and callCount was a total, so slicing the concatenated
list by the total examined the wrong window. With praxis on three requests and
nexus on one, a fourth praxis call landed at index three and paths[4:] never
saw it, while the stale nexus call was reported as new. The mark is now
per server and the paths are taken per server from it.
Neither scenario ever produced an act, so all three fakes saw zero requests and
the fault lever changed no outcome. The two headline capabilities of the
harness had no coverage. act_degraded scripts an act against an enabled
allowlist row and runs it healthy, at 503 and healthy again, asserting the
reply, the call, the absence of a call on a tick, and that nothing was pushed
at him either way. That needed two seams the world did not have: allowlist rows
from the scenario, and a matcher on the real store rather than a nil API, which
would have panicked the moment any scenario produced an act.
expect_no_events compared bus.Len(), which stops growing at the ring capacity,
so a scenario long enough to fill the ring made every later expect_no_events
pass unconditionally. It counts publishes through a subscriber now.
A scenario could not express a fact below full confidence, because write
hardcoded 1.0, and morning_missed annotated its ambient step as if it could.
factPriority branches on exactly that, so no replay could reach the low branch.
signalStep takes a confidence, the ambient step sets the 0.6 the ambient path
writes, and the event line carries the priority so a scenario can assert it.
TestSimulatorIsDeterministic compared the transcript against time.Now, which
fails for the half hour a day the scenario itself covers. It checks that every
stamped line falls inside the scenario span instead. TestSimulatorRefusesBackwardsSteps
tested the forwards case, because reaching the backwards branch ended the test.
A fatalf seam makes the refusal observable.
Smaller notes: the step doc comment now states which assertions are run scoped
and which are step scoped, audioText parses the golden manifest once per world
rather than once per step, and the feminine checks list the masculine form with
its following character, since the earlier check on a comma alone passed on
"записал что ты выпил воды".
Found in review of #79.
Saving a description writes recall corpus. writeNote embeds it under
media:image:<id>, a source no enrollment owns, and the method sits at
AuthRead, so any enrolled module could put a small VLM's guess into what
Maven knows and have it come back in a later turn as something she
believes. The describing half stays a read; save_note is now held to the
same source-scope rule WriteFact is, and the stored text carries a
marker saying it came off a picture.
Three doc comments said the method exists only when vision is enabled
and the code says otherwise. The code is right, and storing without
describing is the state this box is in, so the comments were corrected
rather than the behaviour. A request carrying both data and id used to
take the id branch and drop the bytes without a word; it is refused.
A media dir that cannot be created and a vision endpoint that is a typo
were logged at wiring time and the capability just stayed off, which is
the hardest kind of misconfiguration to notice. Both fail at startup.
runPrune was the one loop started with a bare go and not in the daemon's
WaitGroup, so shutdown did not wait for a prune that was deleting files.
Found in review of #72.
checkPrivate validates the configured endpoint literal and validated
nothing after it. The client followed redirects, so a 302 from the local
llama-server would have sent the photo, as a data URI in the POST body,
to whatever the redirect named. "No provider in this repo may upload a
blob" was true only of the first hop. Redirects are refused now, and the
reply is read through a cap rather than however much the endpoint feels
like sending.
ValidateEndpoint exports the same check so config can fail at startup on
a typo instead of logging once and leaving vision quietly off.
Found in review of #72.
"что нового?" is a greeting, and the feed matcher claimed it: "нового" was a
feed noun and "что" an ask. With no feeds block, which is what ships, the answer
to hello was "я пока не читаю ленты — они не настроены". A newness word now
needs a named topic or a real feed noun beside it. The topic prepositions lose
"о" for the same class of reason: one rune of filler produced a category of
whatever followed it, and then "по этой теме в лентах пока ничего".
An undated feed was re-noted in full on every boot. Dated items are deduped
against the durable mark, undated ones against a map that dies with the process,
so five items became five more on the next start, stamped now, at the top of the
recent-notes window. A crash loop made that a flood. The mark is now set for an
undated feed too, and its existence marks the first poll after a restart as a
resync: those items are recorded as seen rather than written.
A burst larger than max_items lost its middle. The poll walked the feed
newest-first, stopped at the cap, and marked the newest item written, which put
everything below the cap behind the mark forever. The cap now applies to the
oldest candidates and the mark follows what was written, so max_items paces
instead of dropping.
The category tag was read out loud: "Заголовок [технологии]" went through piper
brackets and all, because the answer path took the whole first line. The tag is
parsed off for reading and is now the only thing a topic is matched against.
Matching the whole note meant "что нового про погоду" hit any tech headline
whose link contained "pogod".
Also: the charset comment on dec.Strict described something Strict does not do,
and a skipped feed is named in the log.
Found in review of #66.
Put wrote the blob and then the sidecar. A full disk or a crash between
the two left bytes on disk with no sidecar, and List walks sidecars, so
Prune could never see them: Put returned an error and an image nobody
knew about became permanent. The sidecar goes first, a failed write is
rolled back, and Prune also collects blob files that have no readable
sidecar and are past retention, which picks up whatever an older build
leaked.
The per-blob cap bounds one call and nothing bounded their sum. Content
addressing does not help, because one flipped pixel is a different
digest, so 64 MiB per call and an unlimited number of calls fills the
disk mavend's database lives on. The store now carries a whole-store
budget, seeded from disk at open so a restart does not begin at zero.
Found in review of #72.
The only cap was 64 MiB of input, and a decode bomb is a small file. A
20000x20000 PNG of flat colour compresses to a few hundred kilobytes,
decodes to 400 million pixels, and flattenAndScale then allocated a
second buffer of the same dimensions before scaling anything. That is
3.2 GB of live heap from one request, on a laptop, in the process that
owns the database and the socket, and max_dim never got a chance to
help. The header is read first now and a source over forty megapixels is
refused. The scaler reads the source through At and allocates only the
destination, so flattening no longer doubles the peak.
Found in review of #72.
The fix pass ran in git worktrees, which need models/ symlinked in from the
main checkout to build. .gitignore covered the embedder and llm symlinks but
not stt and tts, so those two were committed as absolute-path symlinks.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TrVSBKe3RFDF4fGYKWYQnX
kami
claude