Compare commits

...

9 Commits

Author SHA1 Message Date
kami ccc5cba2a3 Merge the example-led prompt finding 2026-07-31 15:41:01 +04:00
kami 89d83c0b11 Record the example-led nudge prompt experiment (#393) — it made things worse
Tried rewriting the nudge prompt to lead with five on-topic examples instead
of rules. Three eval runs each side: before 12/13/14 of 15, after 11/12/11.
The loss is all in the address check — formal "вы" and plural imperatives came
back once the "говоришь на ты" rule stopped being its own sentence, and the
on-topic examples leaked their wording into the wrong cases.

Prompt reverted. Only the finding is committed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
2026-07-31 15:40:18 +04:00
kami a97f554802 Merge the informal address prompt rule 2026-07-31 14:54:20 +04:00
kami f4de2fc5e1 Don't let a verb count as the person being talked about
The third-person check asks whether anyone else was named before "он".
A nudge is mostly verbs, and they were counted as possible people, so
"попробуй встать и отдохнуть — у него есть перерыв" passed. Infinitives
and imperatives now join past tense as words that cannot be a person.

A plain noun before the pronoun still blinds it. That needs a parser,
and the comment says so.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
2026-07-31 14:54:20 +04:00
kami eef5d4da4f Tell the phraser to speak to him informally, singular
The prompts stated the feminine self-reference rule but never said whom she is
speaking to, so the model produced formal plural ("Жду вас") and talked about
him in third person ("Он не ел 11 дней"). Adds the address rule right next to
the feminine one, in the nudge prompt and the confirmation prompt.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
2026-07-31 14:52:23 +04:00
kami 09f1696fce Merge the eval label and kill script fixes 2026-07-31 14:32:45 +04:00
kami 80f7322294 Don't fail when docker confirms nothing is running
"Nothing on the host" meant two different things and the script treated
them the same. If docker answers and names no running containers, Maven
really is down and the script should say so and exit 0. Only when docker
cannot be asked is the answer unknown, and that is the case that must
fail loudly.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
2026-07-31 14:32:45 +04:00
kami 02e8786695 Stop the containers instead of claiming success (#380)
docker-compose.yml has no 'pid: host', so each container has its own PID
namespace and pkill on the host matches nothing inside them. The script
then printed "All services gracefully stopped" while mavend, its
llama-server and the rest were still running.

Now it checks for running compose containers first and stops them with
docker compose. If it cannot ask docker and finds nothing to kill on the
host, or anything survives the kill, it says so and exits non-zero
instead of claiming success. The bare-metal path is unchanged apart from
verifying the SIGKILL actually worked, and no longer risks killing the
shell it was launched from.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
2026-07-31 14:30:51 +04:00
kami a788ca3915 Label eval runs with the model the server actually loaded (#379)
The phrasing eval printed "llm (0.8B, ...)" no matter which gguf
llama-server had loaded, so two runs of two different models came out
named the same and were easy to mix up when comparing.

It now asks llama-server over /v1/models, same as the router eval
already did. The helper moved to internal/llm so both share it, and it
now errors instead of returning a blank name when the id field is
missing — an unreachable server gets labelled "unknown-model", never a
plausible-looking guess.

Both eval paths stay opt-in behind MAVEN_LLM_URL; no server needed for
go test.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CGeSZxh1DCtRxmFVSYVGvJ
2026-07-31 14:25:28 +04:00
10 changed files with 255 additions and 22 deletions
+29
View File
@@ -103,6 +103,35 @@ but a large part of the jump is that failure now degrades into Russian instead o
The two remaining failures: one `"..."` recurrence (`routine-stretch`) and one meal nudge
that never says food.
## Tried and reverted: an example-led nudge prompt (#393)
The idea was that a 0.8B copies examples better than it follows rules, so the nudge prompt
was rewritten to lead with five on-topic examples (water, break, pills, morning, service) and
the prose rules were compressed to pay for the tokens: 1190 chars down to 986.
It measured **worse**, three runs each side, same llama-server, same fixture:
| run | before | after |
|---|---|---|
| 1 | 12/15 (address 14) | 11/15 (address 13) |
| 2 | 13/15 (address 15) | 12/15 (address 15) |
| 3 | 14/15 (address 15) | 11/15 (address 12) |
`feminine` and `hisgender` were 15/15 on all six runs, so they measure nothing here. The
regression is all in `address`: 44/45 before, 40/45 after. Formal "вы"/"ваше" and plural
imperatives came back, and so did `"..."`.
Two likely causes, both about the same thing — **examples do not carry a prohibition**. The
old prompt spent a whole sentence on «говоришь на "ты", в единственном числе»; the new one
demoted that to one item in a long "никогда" list, and the model stopped obeying it. And
making the examples on-topic let their *wording* leak: a break case came back as
«Вы давно не пили воду. Выпей стакан.» — the water example, verbatim, in the wrong slot.
That is exactly the failure the laundry/laptop examples were chosen to avoid.
Change reverted. What survives is the measurement: a rule the model must obey needs its own
sentence, and examples must stay off-topic. Also note the before side alone spans 1214 of
15 — this fixture cannot resolve anything smaller than about three cases.
## Broken, found, not fixed
1. ~~**`checkFeminine` only catches half the constraint.**~~ **Fixed** (#381). It scanned for
+1 -1
View File
@@ -29,7 +29,7 @@ func newLLMReplier(c completer) *llmReplier {
return &llmReplier{c: c, stub: voice.NewStubReplier()}
}
const replySystem = `Ты — Maven, домашняя ассистентка (о себе — в женском роде). Подтверди действие РОВНО ОДНИМ коротким предложением (≤120 символов), тепло и по-русски. Не задавай вопросов, не повторяй слова, не добавляй ничего после точки. Отвечай ТОЛЬКО одним объектом JSON с полями "response" (текст) и "mood" (ровно одно из: neutral, happy, thinking, tired, confused).
const replySystem = `Ты — Maven, домашняя ассистентка (о себе — в женском роде). Владелец — мужчина, говоришь с ним на "ты", в единственном числе; никогда не "вы"/"ваш" и не "он"/"его". Подтверди действие РОВНО ОДНИМ коротким предложением (≤120 символов), тепло и по-русски. Не задавай вопросов, не повторяй слова, не добавляй ничего после точки. Отвечай ТОЛЬКО одним объектом JSON с полями "response" (текст) и "mood" (ровно одно из: neutral, happy, thinking, tired, confused).
Пример: {"response": "Записала, что ты выпил стакан воды.", "mood": "neutral"}
Никогда не пиши "..." в поле response.`
@@ -1,4 +1,4 @@
package eval
package llm
import (
"context"
@@ -6,8 +6,20 @@ import (
"fmt"
"net/http"
"strings"
"time"
)
// UnknownModel is the label to print when the server would not say what it has
// loaded. Deliberately ugly: an honest "unknown" is fine, a plausible-looking
// but wrong model name is the bug this whole file exists to prevent.
const UnknownModel = "unknown-model"
// llama-server is local, so never send this through a proxy: this box's
// http_proxy answers 503 for loopback, which would look like "server won't say
// which model it has" when the server is right there and fine.
// A Transport with no Proxy set bypasses http_proxy entirely.
var modelHTTP = &http.Client{Timeout: 10 * time.Second, Transport: &http.Transport{}}
// ModelID asks llama-server which model it has loaded, so a scoring run can
// label itself. Without this a bake-off between two models produces two tables
// that look identical, and the operator has to remember which server was up.
@@ -19,7 +31,7 @@ func ModelID(ctx context.Context, base string) (string, error) {
if err != nil {
return "", err
}
resp, err := http.DefaultClient.Do(req)
resp, err := modelHTTP.Do(req)
if err != nil {
return "", err
}
@@ -38,14 +50,21 @@ func ModelID(ctx context.Context, base string) (string, error) {
if len(out.Data) == 0 {
return "", fmt.Errorf("models: empty list")
}
return shortModelID(out.Data[0].ID), nil
short := shortModelID(out.Data[0].ID)
if short == "" {
// Server answered but the id field was missing or blank. Say so
// instead of handing back an empty label that reads as a real name.
return "", fmt.Errorf("models: no id in response")
}
return short, nil
}
// shortModelID trims the path and the .gguf suffix — llama-server reports the
// file name it was started with, which is too long for a table header.
func shortModelID(id string) string {
id = strings.TrimSpace(id)
if i := strings.LastIndexAny(id, "/\\"); i >= 0 {
id = id[i+1:]
}
return strings.TrimSuffix(id, ".gguf")
return strings.TrimSpace(strings.TrimSuffix(id, ".gguf"))
}
+64
View File
@@ -0,0 +1,64 @@
package llm
import (
"context"
"net/http"
"net/http/httptest"
"testing"
)
// The point of these tests: a wrong-but-plausible model label is the bug, so
// every path that cannot learn the real name must return an error instead of a
// guess. No llama-server needed — a stub server stands in.
func TestModelID(t *testing.T) {
cases := []struct {
name string
body string
code int
want string // "" ⇒ expect an error
}{
{"full path", `{"data":[{"id":"/mnt/hdd1/llms/qwen3.5/Qwen3.5-0.8B.Q4_K_M.gguf"}]}`, 200, "Qwen3.5-0.8B.Q4_K_M"},
{"bare name", `{"data":[{"id":"LFM2.5-1.2B"}]}`, 200, "LFM2.5-1.2B"},
{"empty list", `{"data":[]}`, 200, ""},
{"id missing", `{"data":[{}]}`, 200, ""},
{"id blank", `{"data":[{"id":" "}]}`, 200, ""},
{"server error", `nope`, 500, ""},
{"not json", `<html>`, 200, ""},
}
for _, c := range cases {
t.Run(c.name, func(t *testing.T) {
srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
if r.URL.Path != "/v1/models" {
t.Errorf("asked for %s, want /v1/models", r.URL.Path)
}
w.WriteHeader(c.code)
_, _ = w.Write([]byte(c.body))
}))
defer srv.Close()
got, err := ModelID(context.Background(), srv.URL+"/")
if c.want == "" {
if err == nil {
t.Fatalf("want an error, got label %q", got)
}
return
}
if err != nil {
t.Fatalf("ModelID: %v", err)
}
if got != c.want {
t.Errorf("got %q, want %q", got, c.want)
}
})
}
}
func TestModelIDUnreachable(t *testing.T) {
srv := httptest.NewServer(http.HandlerFunc(func(http.ResponseWriter, *http.Request) {}))
url := srv.URL
srv.Close() // nothing listening now
if got, err := ModelID(context.Background(), url); err == nil {
t.Fatalf("want an error from a dead server, got label %q", got)
}
}
@@ -24,3 +24,19 @@ func TestAddressTimeWordDoesNotBlind(t *testing.T) {
}
}
}
// TestAddressVerbIsNotAnAntecedent — a nudge is mostly verbs, and a verb is
// never who "он" refers to. This exact string passed the check before.
func TestAddressVerbIsNotAnAntecedent(t *testing.T) {
s := "попробуй встать и отдохнуть — у него есть перерыв"
if r := checkAddress(s); r.Pass {
t.Errorf("checkAddress(%q) passed, want a third-person failure", s)
}
// Still missed, and this is the documented hole: "выпей воды, он не пил" has
// a real noun ("воды") before the pronoun, so the scan believes somebody
// else was named. Telling that apart needs a parser, not a suffix rule.
// A named third party still wins over the verbs around it.
if r := checkAddress("сервис упал, он не отвечает"); !r.Pass {
t.Errorf("checkAddress on a real third party failed: %s", r.Detail)
}
}
+22 -8
View File
@@ -340,7 +340,9 @@ func prevWord(words []string, i int) string {
// - it only looks BACKWARD. "Он не отвечает, сервис упал" names the subject
// after the pronoun and is flagged wrongly.
// - any noun earlier in the message counts as an antecedent, even when it is
// not one ("после обеда он не ел" reads as legitimate and is missed). The
// not one ("после обеда он не ел", "выпей воды, он не пил" — both missed).
// Verbs and time words no longer count, which covers the usual nudge, but a
// plain noun before the pronoun still blinds it. The
// common time words are stoplisted so the usual nudge opening does not
// blind it, but a message with any other noun in front still slips through.
// This is the check's real hole; widening it further would start flagging
@@ -407,14 +409,26 @@ var notAnAntecedent = map[string]bool{
"твой": true, "твоя": true, "твоё": true, "твое": true, "твои": true, "твою": true,
}
// looksPastVerb — a past-tense verb needs a subject of its own, so it is not an
// antecedent either. Keeps "сервис упал, он не отвечает" working off "сервис".
func looksPastVerb(w string) bool {
if len([]rune(w)) < 3 {
// looksVerb — a verb is never the thing "он" refers to, so it must not count as
// an antecedent. Past tense keeps "сервис упал, он не отвечает" working off
// "сервис"; the infinitive and imperative endings are here because a nudge is
// mostly made of them ("попробуй встать и отдохнуть — у него есть перерыв"
// slipped through with "попробуй" taken for the person being talked about).
func looksVerb(w string) bool {
r := []rune(w)
if len(r) < 3 {
return false
}
return strings.HasSuffix(w, "л") || strings.HasSuffix(w, "ла") ||
strings.HasSuffix(w, о") || strings.HasSuffix(w, "ли")
for _, suf := range []string{
"л", а", "ло", "ли", // past tense
"ть", "ться", "ти", "чь", // infinitive
"й", "йся", "йте", // imperative
} {
if strings.HasSuffix(w, suf) {
return true
}
}
return false
}
func checkAddress(body string) Result {
@@ -441,7 +455,7 @@ func checkAddress(body string) Result {
if !unicode.Is(unicode.Cyrillic, []rune(p)[0]) && !isLatinWord(p) {
continue // punctuation
}
if notAnAntecedent[p] || prepositions[p] || thirdPersonHim[p] || looksPastVerb(p) {
if notAnAntecedent[p] || prepositions[p] || thirdPersonHim[p] || looksVerb(p) {
continue
}
named = true
+15 -1
View File
@@ -7,6 +7,7 @@ import (
"testing"
"time"
"github.com/kami/maven/internal/llm"
"github.com/kami/maven/internal/phraser"
)
@@ -32,6 +33,7 @@ func TestLLMPhrasingBaseline(t *testing.T) {
// case as a phrasing error and read as "the model cannot phrase".
noProxyLoopback(t)
ctx := context.Background()
f, err := Load()
if err != nil {
t.Fatalf("Load: %v", err)
@@ -44,7 +46,19 @@ func TestLLMPhrasingBaseline(t *testing.T) {
p := phraser.NewLLMPhraserAt(base, cfg)
defer p.Close()
rep, err := Score(context.Background(), "llm (0.8B, built-in persona)", p, f)
// Label the run with whatever gguf the server actually has loaded. It used
// to say "0.8B" no matter what, so two runs of two different models came
// out named the same and were easy to mix up when comparing.
model, err := llm.ModelID(ctx, base)
if err != nil {
// An unlabelled score is still a score, but say so loudly — a made-up
// name in a bake-off table is worse than no name.
t.Logf("could not read model id from %s: %v — report will say %q", base, err, llm.UnknownModel)
model = llm.UnknownModel
}
t.Logf("scoring model %s at %s", model, base)
rep, err := Score(ctx, "llm ("+model+", built-in persona)", p, f)
if err != nil {
t.Fatalf("Score: %v", err)
}
+3 -1
View File
@@ -439,8 +439,10 @@ func (p *LLMPhraser) chatWithSystem(ctx context.Context, system, user string, ma
// as "..." before this. See PHRASING-EVAL-31-07-2026.md.
//
// Russian only, feminine self-reference, second person masculine (the owner is
// a man). One short sentence — the nudge is spoken aloud.
// a man). She talks TO him, informally, singular — never "вы", never "он".
// One short sentence — the nudge is spoken aloud.
const nudgeSystem = `Ты — Maven, домашняя ассистентка. О себе говоришь в женском роде ("я проверила", "я записала"). Владелец — мужчина, обращайся к нему в мужском роде ("ты пил", "ты забыл").
Говоришь с ним на "ты", в единственном числе ("выпей", "встань"). Никогда не "вы"/"вас"/"ваш" и никогда "он"/"его" — ты говоришь ему, а не о нём.
Пиши ОДНО короткое напоминание по-русски: не больше 120 символов и не больше 16 слов. Только по делу.
+3 -3
View File
@@ -61,12 +61,12 @@ func TestLLMRouterBaseline(t *testing.T) {
}
ctx := context.Background()
model, err := ModelID(ctx, base)
model, err := llm.ModelID(ctx, base)
if err != nil {
// Not fatal: an unlabelled score is still a score. But say so loudly,
// because an unlabelled row in a bake-off table is worthless.
t.Logf("could not read model id from %s: %v — reports will say %q", base, err, "unknown-model")
model = "unknown-model"
t.Logf("could not read model id from %s: %v — reports will say %q", base, err, llm.UnknownModel)
model = llm.UnknownModel
}
t.Logf("scoring model %s at %s", model, base)
lr := router.NewLLMRouter(client)
+79 -4
View File
@@ -1,8 +1,10 @@
#!/usr/bin/env bash
# Unified script to stop all Maven services.
# Usage: ./kill-maven.sh
# - Graceful SIGTERM is attempted first.
# - If any process lingers, force with SIGKILL.
# - Docker deploy: `docker compose stop` (see why below).
# - Bare-metal / dev run: graceful SIGTERM first, SIGKILL if anything lingers.
# Exits non-zero if it cannot confirm everything is stopped. It must never say
# "stopped" unless it checked.
set -euo pipefail
@@ -24,6 +26,73 @@ else
LLM='llama-server.*\.gguf'
fi
COMPOSE_FILE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)/docker-compose.yml"
# --- containerised deploy ------------------------------------------------
# docker-compose.yml does not set `pid: host`, so each container has its own
# PID namespace: pkill on the host sees nothing inside them. This script used
# to print "all stopped" while every daemon was still happily running. Stop the
# containers through compose instead — that actually reaches them.
#
# running_containers prints the ids of the project's running containers, or
# nothing. Empty output plus a non-zero return means "could not ask docker",
# which is different from "nothing is running" and is handled below.
running_containers() {
docker compose -f "$COMPOSE_FILE" ps -q --status running 2>/dev/null
}
DOCKER_OK=0
CONTAINERS=""
if command -v docker >/dev/null 2>&1 && [ -f "$COMPOSE_FILE" ]; then
if CONTAINERS="$(running_containers)"; then
DOCKER_OK=1
fi
fi
if [ "$DOCKER_OK" = 1 ] && [ -n "$CONTAINERS" ]; then
echo "--- Maven is running in containers: stopping via docker compose ---"
if ! docker compose -f "$COMPOSE_FILE" stop; then
echo "ERROR: 'docker compose stop' failed. Containers may still be running." >&2
exit 1
fi
echo "--- Verifying containers are gone ---"
LEFT="$(running_containers || true)"
if [ -n "$LEFT" ]; then
echo "ERROR: containers still running after stop:" >&2
docker compose -f "$COMPOSE_FILE" ps >&2 || true
exit 1
fi
echo "All containers stopped."
exit 0
fi
# --- bare-metal / dev run -----------------------------------------------
# pgrep -f matches whole command lines, so a shell that merely mentions
# "mavend" (this script's own parent, for one) shows up. Drop ourselves and our
# parent, otherwise the SIGKILL sweep can take out the terminal you ran this in.
host_pids() {
pgrep -f "$PAT|$LLM" | grep -v -e "^$$\$" -e "^$PPID\$" | paste -sd, - || true
}
HOST_PIDS=$(host_pids)
if [ -z "$HOST_PIDS" ]; then
# Nothing on the host. Whether that means "already down" depends on whether
# we managed to ask docker, and the two must not read the same.
if [ "$DOCKER_OK" = 1 ]; then
# Docker answered and named no running containers, and there is nothing
# on the host either. That is a real answer: Maven is already stopped.
echo "Nothing to stop: no Maven processes and no running containers."
exit 0
fi
# We could not ask docker, so Maven may be alive in a container we cannot
# see. Saying "stopped" here is the exact false success this script had.
echo "ERROR: no Maven processes on this host, and docker could not be asked." >&2
echo " If this is the container deploy it may still be running:" >&2
echo " docker compose -f $COMPOSE_FILE stop" >&2
echo " Nothing was stopped. Check by hand before assuming Maven is down." >&2
exit 1
fi
echo "--- Sending graceful SIGTERM to Maven services ---"
pkill -TERM -f "$PAT" || true
# mavend's Pdeathsig SIGKILLs its llama-server on exit, but sweep strays too
@@ -32,12 +101,18 @@ pkill -TERM -f "$LLM" || true
echo "--- Verifying processes are gone ---"
sleep 1
PIDS=$(pgrep -d ',' -f "$PAT|$LLM") || PIDS=""
PIDS=$(host_pids)
if [ -n "$PIDS" ]; then
echo "Warning: some processes still alive. PIDs: $PIDS"
echo "--- Force killing with SIGKILL ---"
echo "$PIDS" | tr ',' '\n' | xargs -r kill -9
sleep 1
LEFT=$(host_pids)
if [ -n "$LEFT" ]; then
echo "ERROR: still alive after SIGKILL. PIDs: $LEFT" >&2
exit 1
fi
echo "Done (SIGKILL)."
else
echo "All services gracefully stopped."
fi
fi