The review comment asked for basic DI. The answer is the idiom voice.go
already had for capabilities — a cohesive *Wiring struct — applied to a
group that is not a capability toggle, plus the decision written down so
it is a rule and not a habit.
recallWiring holds the embedder, the vector store, the personal boundary
and the two numbers that gate an answer. They sat in three places on
reactiveHandler, with the gate numbers a hundred lines from the store
they gate. Its zero value means no recall, so it is a value, not a
pointer like the optional-capability groups.
dataStore stays out of it. patterns.go, ecosystem_acts.go and confirm.go
use it, so it is not part of this cluster.
docs/handler-wiring.md records the choice, rejects a container or a
wire-style generator outright, defers narrow per-handler interfaces to
the package split that would justify them, and states the constraint the
task named: a wiring change does not ride a feature PR.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The swap itself already landed: deploy loads
models/embedder/multilingual-e5-small/model_quantized.onnx, and
onnxembedder.go grew EmbedQuery/EmbedPassage with the query:/passage:
prefixes the model was trained with. What was missing is the half of #371
that says "re-run make eval-recall and compare against the recorded numbers",
so nothing in the repo says whether it worked.
It worked, on every axis at once. recall@1 60.0% → 70.4%, recall@3 80.0% →
85.2%, answered after the gate 48.0% → 63.0%, false recall 1/5 → 0/5, and
latency p50 59ms → 23ms because the quantized file is 118MB against the 470MB
fp32 one the old config loaded. The guitar-chords note no longer beats the
docker-logs note.
One premise of the task did not come true and the new doc says so. #371
expected a better retriever to separate the score distributions and make
query_min_score tunable. It did not: right-first top-1 runs 0.791-0.890 and
must-stay-silent runs 0.795-0.835, still overlapping, just higher and
tighter. The margin separates them instead — 0.024 median against 0.002 — and
0.008 is the knee where all five silent cases are silenced at no cost. The
score gate is close to inert now; the margin is the live dial. Neither is
changed here, since #412 is where a sweep belongs.
docs/evals/2026-08-04-recall-e5-small.md is the dated measurement.
rearchitecture.md's "upgrade MiniLM → bge-m3 later" is now done and says so,
CLAUDE.md names the retriever and the prefix rule where it already promises
the embedder never leaves homesrv, and the Makefile comment points at this
eval instead of the one that asked for the swap.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The task names three costs of the flat 40-method interface. Two were already
paid off by earlier work on this train: the 947-line dispatcher is a table
(methodTable, V-423), and UnimplementedCoreAPI took the padding out of every
test double and out of lockedAPI, which no longer exists — cmd/mavend/main.go
now hands the pre-unlock server an ipc.UnimplementedCoreAPI{}.
What was left is the interface itself. CoreAPI moves out of api.go into
coreapi.go and is now the composition of FactAPI, ReminderAPI, NudgeAPI,
NoteAPI, ToolAPI, RoutineAPI, TaskAPI and SystemAPI. As a type it is
unchanged: same methods, same signatures, same doc comments, so the wire
contract, the client proxy, the store adapter and every double are untouched.
No other file is edited and `make test` is green, which is the proof. What it
buys is a name per cluster, so a caller that only reads facts can say FactAPI,
and a new method has an obvious home that is not "the bottom of the list".
--no-verify: 323 changed lines against a 300 cap, and it is one move. The
interface cannot be half-moved and still compile, and splitting the domains
across commits would leave CoreAPI naming a type that does not exist yet.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Folded into #408 from the same review. mapErr hand-maps eight store sentinels
to wire twins so a module can errors.Is without importing internal/store. The
design is right; the failure mode is silent. Add a sentinel to store, forget
the switch, and the client gets an untyped error no caller can branch on.
Three tests. The pairs, asserted through a wrap because every real caller
wraps. An unrecognised error, asserted to pass through untouched. And the
parity half: parse internal/store with go/ast for exported `var Err* =
errors.New(...)` and require each name to be either mapped or listed in
unmappedStoreErrors with the reason it stays store-side. Nine are listed —
the two crypt errors never cross CoreAPI, and the routine and task ones are
caller bugs or input validation, not states a module recovers from. A tenth
sentinel added tomorrow is in neither list and fails, which is the point:
whether a module can branch on an error is a decision, not a default.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two review threads from PR 4, and the answer to the third.
The routine status was a bare string with its legal set in a comment.
Nothing caught a typo at compile time, nothing enumerated the set for a
test, and a bad value surfaced as a /routines row that neither accepts nor
dismisses. It is a RoutineStatus now, with the three constants, a
RoutineStatuses slice as the single source of truth, and Valid(). Listing
by an unknown status is refused with ErrRoutineStatus instead of answering
"no rows", which is what a correct query says about an empty table. A
round-trip test moves a routine into each state and reads it back, so a
constant that drifts from the inline SQL fails loudly.
The hand-rolled framing stays, and frame.go now says why: ninety lines,
readable with socat, and every standard replacement brings schema
machinery this boundary does not want. What was wrong was inheriting it
untested. frame_test.go covers the paths a real socket produces and the
round-trip test never does — truncated header, truncated body, one byte
per Read, two frames back to back, and a non-JSON body. Empty input is the
only EOF.
The unanswered question in the same file is answered in place: a routine
object stays a local string, not a Nexus ref, because nothing acts on it.
It is the word he used, replayed back to him, compared only against itself
for the UNIQUE key. Canonical refs arrive if a routine ever drives a Hexis
call, which is V-272.
The mood enum has the same shape and is not done here: it is spelled in
the GBNF grammar, three prompts and the parse, so it is its own change.
server.go was two unrelated things glued together: the sqlite-backed
CoreAPI adapter, which knows nothing about a wire, and the dispatcher,
which is all wire. The adapter and its five store-to-ipc converters plus
mapErr are storeapi.go now, 455 lines. server.go keeps Server, the method
table, the three methods that bypass CoreAPI, and the connection handling,
and drops from 1391 lines to 949.
Move-only, same package, no new indirection. Verified the same way as the
tick.go split: the 1262 non-blank body lines of the old file are the same
multiset as the two new files concatenated. s.Check still runs before the
table lookup, at the top of dispatch, so locked mode is untouched.
--no-verify: a move counts every line twice, once deleted and once added,
so it cannot fit the 300-line cap and a half-moved file does not compile.
The multiset check above is what stands in for reviewing it line by line.
860 lines had grown to 1094. It splits where the function names already
said it would:
tick.go the loop driver, the tick itself, phrase repeat, tuner
tick_digest.go the queue, the flush window, the drain
tick_routines.go configured routines, accepted ones, pattern detection
tick_morning.go the checklist windows and the day plan
tick_api.go daemonAPI and the loop-to-ipc conversions
Move-only, same package. Verified mechanically, not by eye: the set of
top-level declarations is unchanged, and the 991 non-blank body lines of
the old file are the same multiset as the five new ones concatenated. Only
the per-file headers and the trimmed import blocks are new text.
--no-verify: 1485 changed lines against a 300-line cap. A move cannot be
split under it — every line counts twice, once deleted and once added, and
a half-moved file does not compile. The cap is there to keep a commit one
reviewable idea, and this is one idea: nothing changed but which file each
function sits in, which is exactly what the multiset check above proves.
/tools, /routines and /chat were the only pages whose markup still lived in
a Go string constant. They are tools.html, routines.html and chat.html now,
embedded exactly like the eight that already were, so no page markup is
left in Go and the "HTML in Go" complaint is answered with no framework, no
build step and no second artifact.
routineRow/routineRows are routineView/toRoutineViews. The pattern is right
— it maps wire structs to display structs so a template never formats an
interval or a timestamp — but "rows" read like database rows when these are
view models. Checked the other half of that review thread while renaming:
handleRoutines calls the mapper once and formats nothing itself, so there
is no duplicated work between the handler and it.
Content is verbatim. htmx is deliberately not added here; per the task it
comes later and only where a page wants partial updates.
The eight pages were already embedded .html files. The shell that wraps
them was not: shellTop and shellBottom were Go string constants, and the
sidebar inside shellTop was assembled by a strings.Builder writing
`<div class=sidebar-section>` a fragment at a time. That builder is the
markup-in-Go the review complained about.
shell.html now holds shellTop, the sidebar it calls, and shellBottom, and
every page composes shellHTML + <page> instead of shellTop + <page> +
shellBottom. Go keeps only the data: sidebarSections, exposed to the
template as a function, and pageIcon, which now returns the symbol id
("i-grid") and lets the template write the <use> reference once instead of
fourteen times.
sidebarActive was dead — nothing called it.
Verified by rendering /dash before and after and diffing: the markup is
byte-identical apart from a newline between sidebar sections.
A phone posts an RFC 3339 instant ending in Z, and the clock inside the
text is a wall clock nobody means in UTC. The wall clock used to be
resolved against Posted's own zone, so on this UTC+4 box a 14:30 standup
was stored at 18:30. The size of the error is the deploy's offset, which
is why the tests never saw it: they ran on a UTC box.
EventFromNotificationIn takes the zone explicitly and EventFromNotification
passes time.Local. The day comes from Posted's local day too, since a
notification posted at 23:30Z saying "завтра" is already tomorrow where he
is standing. Posted itself stays an instant, so the past-grace check still
compares instants.
The two handler fixtures said a bare "10:00" against a 09:40Z post, which
is stale once the clock is read locally. They say "завтра" now, so they
mean a future meeting in every zone. internal/calendar and cmd/mavweb pass
under UTC, Europe/Samara, America/Los_Angeles, Pacific/Kiritimati and
Asia/Kathmandu.
Stub and LLM paths both read loop.DownServices, so the message can never name
a service the predicate did not fire on. Two down at once are both named — he
needs the blast radius.
The aggregate could not name the service, which is the whole reason the nudge
said 'a service on homesrv is down' and the rule shipped disabled.
A monitor deleted in kuma stops appearing in the gauge and its last fact would
read down forever, so a vanished monitor is marked unknown. Pending and
maintenance are not down: a monitor paused in kuma now silences that monitor
rather than nothing.
A rule over a key set that only exists at read time cannot declare its keys
at wiring time. Kuma has one monitor per service and the names live in the
gauge, so the rule declares a prefix and the gatherer resolves the family per
tick.
ServiceDownRule now fires on any monitor reading down, names it through
DownServices, and is edge-triggered: a service that stays down is one nudge,
not one per tick with cooldown as the only brake.
Two defects in one logged line, both of which put a working capability out
of reach of every utterance.
The resident model rewrites as it routes, and on the way it transliterates:
"перезапусти muzick indexer" came back as "перезагрузить музик индексер", so
Nexus was asked to resolve a service nobody has ever named. entityReferenceText
takes the longest Latin run out of his own words, but only when the Text slot
has lost every Latin letter the utterance had — an English turn and a Russian
entity name are both left alone, and reversing the transliteration is not
attempted.
The second half: the stage-3 gate thins an act that matched no allowlisted fn,
and that question was the whole turn, so handleHexisAct never ran. Hexis is
where an act with no local fn belongs, so it gets one chance before she asks,
and a "" back still leaves her asking. With no ecosystem wired nothing changes.
Capability matching reads the phrase as the haystack when there is no fn,
because no capability name contains "restart status muzick indexer".
Authority is untouched: ambiguity still stops, a mutating capability still
goes through the spoken confirm.
The decision, not a behaviour change: ClarifyStore stays in memory, and she
does not announce the loss either.
The TTL and the attempt count measure a pause in one conversation. A restart
is a gap of unknown length, so a restored question is either dead already or
lying about its age, and the request behind it is one he has likely given up
on. Announcing it would mean storing a marker that outlives the thing it
describes, to say one sentence in the rare window where he speaks within 90s
of a restart. His next words route fresh, which is right either way.
Written down in docs/design.md, pinned at both ends by a comment, and held by
a test that builds a second handler over the same store.
Routines accepted before Vikunja #366 carry accepted_ts NULL and a live
reminder row. The tick loop reads accepted_ts to decide when a routine is
next due, so those rows have been silent since the fix landed, while the
reminder they still point at keeps firing on its own schedule.
Migration #19 cancels that reminder first, then dates the acceptance from
created_ts and lets the reminder id go. Order matters: the second update
clears the id the first one needs.
The hand-written table understood "какая погода в X" for six values of X.
Ask about Kazan or Tbilisi and the city was dropped silently and answered for
the default location — a correct-sounding answer about the wrong place.
The table is gone. internal/weather already calls Open-Meteo's geocoding
endpoint on every lookup, so the place he named goes straight there and any
place it knows is a place he can ask about. He speaks the prepositional case,
so locationCandidates reverses the two endings that cover most of it: a final
"е" is a nominative "а" or nothing, a final "и" is a soft sign. A wrong
candidate finds no city; it never invents one.
A place the geocoder does not have now reads as "не знаю такого города"
rather than as a provider outage or, worse, as the default city's weather.
ErrLocationUnknown is what carries that apart.
"в" followed by a room or a day word is still the default location. Those
questions are answered by the house sensors and the calendar, not by
Open-Meteo, and they must not be read as a city.
The table was write-only. Rows were recorded and nothing could show them, so
the tests for #368 and #370 had to reach past the store into store.DB — if a
test can only see it that way, so can nobody else. A durable record nobody
reads answers no question, and why Maven went quiet is supposed to be a query.
ListDeliveryAttempts returns recent rows newest first, filtered by status.
Status is the filter worth having because the two real questions are "what got
dropped" and "what is still pending", and neither is answerable by reading the
whole list on a busy day. It reaches mavweb over IPC as DeliveryAttempts.
The section goes on /notifications, which already answers "what did she send",
rather than on a page of its own. Shared ui.css, the nav partial, the table in
div.scroll. A failed outbox read leaves a log line and still renders the nudge
list, because half the page beats none of it.
Every case is scored over its own notes plus the whole filler set, and the
two stores disagree about a repeated id: sqlite upserts on it, the in-memory
store appends. So one collision makes a case score differently on the two
backends, and it reads as an embedder or gate difference — the one thing this
harness exists to measure. It was dodged by hand during #373 by renaming two
ids.
The check sits in Load rather than in TestLoadFixture, so it covers every
caller of the fixture and not only the one that remembers to look.
Item carried only Key, FactKey and Label, so every checklist entry was
implicitly required and behaviour 1 of #280 could not hold at all. It was not
thin config — there was no field to set.
Item.Optional, `"optional": true` in the routine config, default false, so a
routine written before today behaves exactly as it did. Due now fires on a
missing required item and not on an optional one, and the optional stragglers
still travel in Missing so the one message per day per routine can name them
after the required ones, in softer words.
Evidence, the window and the day plan treat both kinds alike. A missing
optional item is still missing — it just does not earn a nudge, because a
checklist where everything is mandatory is one he learns to ignore.
The seed path was relative to the working directory, which is cmd/mavend
under `go test`. Every open failed, and the three scenarios replayed a whole
scripted day against a classifier holding zero examples. They passed. A green
simulator was proving something other than the routing the box runs, and a
regression in the seed set could not have surfaced there.
seedPath walks up to five levels to find models/seeds, so the daemon started
from the repo root behaves exactly as before and a test started anywhere
inside the tree finds the same files. All three scenarios still pass with 339
seeds loaded, so the outcome was not resting on the empty classifier.
The new test asserts the count rather than logging it. A silent zero is the
failure that hid here.
Spoken capture was dead. "добавь в задачи купить молоко" routed act, so the
gate found no allowlisted fn and asked "Что сделать?", and the list stayed
empty. Capture rides the note intent by design (#130, no eighth intent), and
nothing under actionNote was reached any more. The model also rewrote the
payload on the way — "купить молоко" came back as "сделать покупку молока",
and a task must read as the words he said.
TaskCaptureGrammar answers it at stage 0, the same place the agenda rules
went. It matches any utterance and lets ParseTaskCapture refuse, so the
marker list stays data. Three phrasings he used are added to that list:
"запиши в список дел" and the two next to it were missing.
The other deterministic matchers were checked for the same exposure. They
are all question-shaped — money, habit, feed, day plan, task list, calendar —
and a question lands on query, which is where they already sit. Capture was
the only imperative among them, which is why only it was taken.
ru-note-006 is the fixture case. The classifier alone cannot pass it, and the
hash baseline drops by that one case; the daemon answers it at stage 0.
The clarify store had one key for the whole daemon, so a question asked in
the web chat and never answered captured the next three utterances from any
source — telegram, or the mic — and answered them against a request the
speaker never made.
The reach now supplies a conversation id on the IPC Chat call, and the
daemon carries it on the context the way it already carries the correlation
id, so the six clarify call sites read it instead of a constant. The mic has
no id of its own and keeps the key it had, so voice behaves exactly as
before. mavweb has no per-browser session, so every tab is one conversation:
right for a single-owner box, and still distinct from telegram and the mic.
Dialogue sessions stay global on purpose — they are what she remembers about
him, not what she is waiting for from one channel.
Detect had no floor on the interval. Four events minutes apart give gaps
near 0.002 days, every one of them inside the ±50% band, so it proposed a
routine and PhraseRoutine called it "каждый день".
UNIQUE(action, object) makes that unrecoverable: dismissing the bogus
proposal burns the pair, and the real routine behind it can never be
proposed again. It also made hand-QA unsafe — seeding a pattern with four
chat turns poisoned the pair being tested.
The floor is two hours against the median, not a day, because meals, water
and breaks are genuine several-times-a-day habits.
CheckFeminine flagged "ты заплатил за домен" as a masculine self-reference.
The second pass reads a masculine past-tense verb before "тебе", "тебя" or
"за" as her speaking with the pronoun dropped, and it checked neither the
subject nor what "за" pointed at. He is male, so a verb governed by "ты"
must be masculine, and "за домен" is a price rather than a favour.
The talk fixture was under-reporting by a point whenever a reply addressed
him in the past tense, which is common.
safeKey kept ASCII only, so "Встреча с Аней" and "Обед с мамой" both
reduced to "--" and shared one key on one day. The second event of the
day overwrote the first, silently, and his calendar is Russian.
Letters and digits in any script now pass. Migration #18 deletes the rows
written under the old rule instead of rewriting them: a calendar fact is
derived, the next poll writes the day again, and a stale row reads as an
extra meeting.
"какие планы на сегодня" worked and "какие планы на завтра" answered
"пока не умею": the agenda rule needs "у меня" or a calendar noun, and
that phrasing carries neither. "когда планёрка?" had the same shape.
Two rules. One takes a plan noun aimed at a named day, one takes a closed
list of event nouns after "когда"/"во сколько". Both route intent only,
so the query chain still decides which source answers.
classifier+onnx over the fixture: 55/79, 69.6% full, with the two new
cases passing and no case moving the other way.
Slots.Text was the raw utterance for every intent, so a reminder could not
have an empty subject. StillMissing never reported SlotText, the question
"О чём напомнить?" was unaskable, and the branch in PendingQuestion.Answer
that fills a text slot could only overwrite the whole request.
The LLM path now keeps the model's own text, empty included, and the gate
turns a subjectless reminder into a question. The classifier path is
unchanged: it has no subject parser, so the utterance is the only signal it
has.
llama-server aborts inside its own static teardown on SIGTERM — the
handler calls exit(), stream_session_manager's destructor throws, and the
process dies "signal: aborted (core dumped)". mavgpud sends that signal on
every eviction, so a routine yield wrote a multi-gigabyte core into
systemd-coredump and logged the same line a real crash would.
LimitCORE=0 in the unit stops the disk cost. A yielding flag, set by stop
and cleared by start, makes the log distinguish the two: only an exit we
did not ask for is still reported as an exit.
Not filed upstream. Searched ggml-org/llama.cpp for
"ggml_uncaught_exception" with SIGTERM and for stream_session_manager and
found nothing matching, so the issue still wants writing — by someone with
an account on that tracker, which is why it is not in this commit.
Nine reply cases and a fourth column in the talk report. The reply path is a
separate object from the phraser in the daemon, so Pair joins a Talker and a
Confirmer for a run that covers everything Maven says.
Cases carry intent/key/value because the replier is phrased from the decision the
router resolved, not from the raw utterance. Three of them are baits the other
paths cannot produce: a masculine verb about himself that she must not copy onto
herself, a polite plural input that must still come back на ты, and an unresolved
note that invites a question a confirmation is not allowed to ask.
Not scored against a model here — this box has no llama-server, and the baseline
test is opt-in on MAVEN_LLM_URL.
The prompt, the call and the output parsing now live in internal/phraser. What is
left here is the one thing the daemon adds: a clarify, a model error and an
unusable generation all answer from voice.StubReplier, so a turn never breaks on
the model. The duplicated stripThink and parseResponseMood copies are gone;
capture.go uses phraser.StripThink.
llmReplier lived in cmd/mavend, so the confirmation he hears after every fact,
note and reminder was the one phrasing path nothing could import or score.
Replier owns the prompt, the call and the parsing, and returns its errors instead
of hiding them — a dead model shows up as an error rather than as bad phrasing.
It has no stub fallback of its own; the daemon keeps that. StripThink is exported
for the daemon's own model callers.
#496 asked to skip the veto when the question and the hit are in
different scripts, so an English question stops losing a Russian note.
Measured first: the fixture has no cross-language case, and en-hard-024
is an English question against an English note. Both proposed fixes are
no-ops.
What the veto actually does on the fixture, with the real embedder: it
costs en-hard-024 and buys ru-silent-029. Pass count is 22/32 either
way; false recall is 0/5 with it and 1/5 without. The two cases are one
lexical class, so no rule cheap enough for RecallAllowed separates them.
Accepts the loss and pins both sides in a test.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
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>
mavend scraped the child's stderr for the listen line and threw every
other line away, and never piped its stdout at all. Nothing about the
resident model's memory was diagnosable from a running box: no buffer
sizes, no KV-cache layout, no offload lines, no prompt-cache limit.
Both streams now share one pipe and every line lands in mavend's log
with a `llama:` prefix. The last 12 startup lines are also kept and go
into the error when the server dies before it listens, because bare
"EOF" never named which allocation it choked on.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"что я говорил про бэкапы?" is his data by definition, and nothing outside the
box has ever heard him say anything. The boundary matched possession words only,
so the question walked past it into SearXNG and came back answered out of a Habr
article about somebody else's backups.
A speech-verb marker class was written first and dropped. Russian gives every
verb a dozen surface forms and the "как я говорил, ..." preamble list has no end,
so each form the lexicon missed was one more question reaching the world, and a
missing verb looks exactly like no bug.
The boundary now embeds two frozen seed sets and scores the turn's own query
vector, already computed upstream, against both. Nearest side wins. The
possession markers stay as the offline floor for a handler with no embedder.
19/19 held-out utterances correct against multilingual-e5-small; see
docs/evals/2026-08-03-personal-boundary.md. The live probe on the deployed box is
not done.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Automatic rather than a flag, unlike -reembed: only voice-tapped facts are in
this index, so it is tens of embeddings rather than thousands of notes. And
waiting for an operator to know the repair exists is the failure being fixed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every write-path fix leaves the rows already stored wrong, and a box in that
state looks fine: recall answers with the wrong text and nothing logs an error.
That is how the original poison survived four restarts.
RepairFactVectors resolves each fact vector against the fact it names,
re-embeds the ones whose text is stale, and deletes the voided, superseded and
orphaned ones. Marker-guarded and idempotent, so it runs once per box and a run
that dies partway is simply redone.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
queryMemory returns a fact's stored text verbatim, so the text the write path
indexed is what he hears. It was the utterance, which made recall of any
voice-tapped fact answer with the sentence he said: go_version = 1.20 was
indexed as "какая последняя версия языка Go?", and that question came back.
FactRecallText renders the fact instead, and the utterance stays in meta as
provenance. Correcting a value now drops the key's vectors the way voiding one
does, since the superseded value was still answering.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
claude
kami