Merge pull request 'Run the persona checks inside the daemon before she speaks' (#143) from task/399-run-the-persona-checks-inside-the-daemon into master

The checks stay in the eval package and the daemon calls three of them:
feminine, address, and a new leaked-reasoning test. No retry — it doubles
the latency on the turn that is already going badly, and on the nudge path
the moment has passed. A failure falls back to the deterministic floor and
is logged with the whole rejected text and counted by check name.

hisgender is deliberately not run: the simulator showed it rejecting
"записала, что ты выпил воды", which is her own correct self-reference.

cmd/mavend/personaguard.go

@@ -0,0 +1,134 @@
+package main
+
+import (
+	"context"
+	"log"
+	"regexp"
+	"strings"
+	"sync"
+
+	"github.com/kami/maven/internal/delivery"
+	"github.com/kami/maven/internal/loop"
+	"github.com/kami/maven/internal/phraser"
+	"github.com/kami/maven/internal/phraser/eval"
+)
+
+// The persona checks, run before she speaks (Vikunja #399).
+//
+// RunChecks and RunTalkChecks only ever ran from the eval package, so
+// everything the fixtures measured was offline knowledge: we could say "about
+// one reply in three is broken" and still ship every one of them. This runs the
+// cheap half of that on the live path, and replaces a failing message with the
+// deterministic floor.
+//
+// Which checks: the unambiguous string tests only — feminine self-reference,
+// how she addresses him, and a leaked-reasoning test. Not length, which is
+// path-specific, and not ontopic, which compares against fragments the fixture
+// supplies and runtime does not have. Not hisgender either — see guardSpoken.
+//
+// No retry. A retry doubles the latency on the exact turn that is already going
+// badly, and on the nudge path the moment has passed.
+//
+// The known cost, written down because it is real: a wrongly flagged good reply
+// is replaced by a flatter stub one. That is the right trade — a stub sentence
+// is dull, a leaked reasoning trace is broken — but it means these checks can
+// no longer be tuned for sensitivity alone.
+
+// checkLeak — the name reported when the model's scaffolding reaches the text.
+const checkLeak = "leak"
+
+// leakPatterns — reasoning and protocol that belongs to the model, not to him.
+// The resident model is a Thinking variant, so an unclosed reasoning block is
+// the failure mode, not a hypothetical (Vikunja #398).
+var leakPatterns = []*regexp.Regexp{
+	regexp.MustCompile(`(?i)<\s*/?\s*think`),
+	regexp.MustCompile(`(?i)thinking\s*(process|:)`),
+	regexp.MustCompile(`(?i)^\s*(assistant|user|system)\s*:`),
+	// Raw contract JSON: the parser already unwraps a good one, so a body that
+	// still carries the keys is one it could not read.
+	regexp.MustCompile(`"(response|mood|body|summary)"\s*:`),
+	// The persona block quoted back at him.
+	regexp.MustCompile(`(?i)(ты\s+—?\s*мэйвен|системный промпт|system prompt)`),
+}
+
+// checkPersonaLeak reports whether the model's own scaffolding is in the text.
+func checkPersonaLeak(body string) (string, bool) {
+	for _, re := range leakPatterns {
+		if m := re.FindString(body); m != "" {
+			return "leaked " + strings.TrimSpace(m), false
+		}
+	}
+	return "", true
+}
+
+// personaRejects counts what the guard caught, by check name, so the real
+// production rate is knowable rather than inferred from the fixture.
+var personaRejects = struct {
+	mu sync.Mutex
+	by map[string]int
+}{by: map[string]int{}}
+
+func personaRejectCounts() map[string]int {
+	personaRejects.mu.Lock()
+	defer personaRejects.mu.Unlock()
+	out := make(map[string]int, len(personaRejects.by))
+	for k, v := range personaRejects.by {
+		out[k] = v
+	}
+	return out
+}
+
+// guardSpoken checks a phrased message. It returns the failed check and false
+// when the message must not be said; path names the caller, for the log.
+//
+// An empty message passes: the caller already treats that as a failure and
+// falls back on its own, and reporting it as a persona breach would put a
+// misleading line in the count.
+func guardSpoken(path, body string) (string, bool) {
+	if strings.TrimSpace(body) == "" {
+		return "", true
+	}
+	if detail, ok := checkPersonaLeak(body); !ok {
+		return rejectSpoken(path, checkLeak, detail, body), false
+	}
+	// Feminine and address only. HisGender is not run here: it reads a
+	// sentence-initial feminine verb with no pronoun — "записала, что ты выпил
+	// воды" — as a woman being addressed, when it is her own correct
+	// self-reference. Offline that is a point of score; on this path it would
+	// replace a good reply with a stub one on every fact she confirms.
+	for _, r := range []eval.Result{eval.Feminine(body), eval.Address(body)} {
+		if !r.Pass {
+			return rejectSpoken(path, r.Name, r.Detail, body), false
+		}
+	}
+	return "", true
+}
+
+// rejectSpoken logs what she nearly said and counts it. The whole text, not a
+// prefix: the point of the log line is that the failure can be read back later
+// and argued with.
+func rejectSpoken(path, check, detail, body string) string {
+	personaRejects.mu.Lock()
+	personaRejects.by[check]++
+	personaRejects.mu.Unlock()
+	log.Printf("persona: %s rejected on %s (%s): %q", path, check, detail, body)
+	return check
+}
+
+// guardNudge checks a phrased nudge and falls back to the deterministic floor
+// when it fails. The nudge path, unlike the reply path, cannot ask again: the
+// tick has already decided she speaks, so the choice is the floor's wording or
+// a broken sentence.
+func guardNudge(pn delivery.PhrasedNudge, cand loop.Candidate) delivery.PhrasedNudge {
+	if _, ok := guardSpoken("nudge", pn.Body); ok {
+		return pn
+	}
+	stub, err := phraser.NewStub().PhraseNudge(context.Background(), cand)
+	if err != nil {
+		// The Stub is templates over the candidate and does not fail. If it
+		// somehow does, the model's text is still what the rule decided to
+		// say, and saying nothing is the worse outcome.
+		return pn
+	}
+	return stub
+}

cmd/mavend/personaguard_test.go

@@ -0,0 +1,75 @@
+package main
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/kami/maven/internal/delivery"
+	"github.com/kami/maven/internal/loop"
+)
+
+func TestGuardPassesWhatSheShouldSay(t *testing.T) {
+	good := []string{
+		"записала: купить хлеб.",
+		"поняла, напомню в 11:00.",
+		"ты не пил воду с утра.",
+		"я рада, что получилось.",
+		"",
+	}
+	for _, body := range good {
+		if check, ok := guardSpoken("test", body); !ok {
+			t.Errorf("guardSpoken(%q) rejected on %s", body, check)
+		}
+	}
+}
+
+func TestGuardStopsWhatSheShouldNot(t *testing.T) {
+	bad := []struct {
+		body string
+		want string
+	}{
+		{"<think>он просил воду</think> попей воды.", checkLeak},
+		{"Thinking Process: он давно не пил.", checkLeak},
+		{`{"response": "попей воды", "mood": "neutral"}`, checkLeak},
+		{"я напомнил тебе про воду.", "feminine"},
+		{"вы давно не пили воду.", "address"},
+	}
+	for _, c := range bad {
+		check, ok := guardSpoken("test", c.body)
+		if ok {
+			t.Errorf("guardSpoken(%q) let it through", c.body)
+			continue
+		}
+		if check != c.want {
+			t.Errorf("guardSpoken(%q) failed on %s; want %s", c.body, check, c.want)
+		}
+	}
+}
+
+func TestGuardCountsWhatItCaught(t *testing.T) {
+	before := personaRejectCounts()[checkLeak]
+	if _, ok := guardSpoken("test", "<think>…"); ok {
+		t.Fatal("a leaked reasoning block was let through")
+	}
+	if after := personaRejectCounts()[checkLeak]; after != before+1 {
+		t.Errorf("leak count %d; want %d", after, before+1)
+	}
+}
+
+// TestGuardNudgeFallsBackToTheFloor — a broken nudge is replaced by the
+// deterministic wording, not dropped and not retried.
+func TestGuardNudgeFallsBackToTheFloor(t *testing.T) {
+	cand := loop.Candidate{Rule: loop.Rule{Name: "water"}}
+	bad := delivery.PhrasedNudge{Candidate: cand, Body: "Thinking Process: он не пил.", Mood: "neutral"}
+	got := guardNudge(bad, cand)
+	if got.Body == bad.Body {
+		t.Fatal("the broken nudge was delivered unchanged")
+	}
+	if strings.TrimSpace(got.Body) == "" {
+		t.Fatal("the nudge was dropped rather than re-worded")
+	}
+	good := delivery.PhrasedNudge{Candidate: cand, Body: "попей воды.", Mood: "neutral"}
+	if guardNudge(good, cand).Body != good.Body {
+		t.Error("a good nudge was replaced")
+	}
+}

cmd/mavend/replier_llm.go

@@ -35,5 +35,10 @@ func (r *llmReplier) Reply(d router.Decision) string {
 	if err != nil || out == "" {
 		return r.stub.Reply(d)
 	}
+	// The persona checks, on the live path (personaguard.go). A reply that
+	// leaks reasoning or calls him "вы" is worse than a flat one.
+	if _, ok := guardSpoken("reply", out); !ok {
+		return r.stub.Reply(d)
+	}
 	return out
 }

cmd/mavend/tick.go

@@ -176,6 +176,9 @@ func (t *tickLoop) tick(ctx context.Context, now time.Time) {
 			t.queueNudge(ctx, cand, state, now)
 		} else {
 			pn, err := t.phraser.PhraseNudge(ctx, *cand)
+			if err == nil {
+				pn = guardNudge(pn, *cand)
+			}
 			if err != nil {
 				log.Printf("tick: phrase nudge %s: %v", cand.Rule.Name, err)
 			} else {

internal/phraser/eval/checks.go

@@ -67,6 +67,19 @@ func RunChecks(c Case, body, mood string) []Result {
 	}
 }
 
+// Feminine, HisGender and Address expose three checks one at a time, so the
+// daemon can run them on a phrased message before he hears it (Vikunja #399).
+// Only these three: they are unambiguous string tests with nothing to compare
+// against, while length is path-specific and ontopic needs the fixture's
+// expected fragments, which do not exist at runtime.
+func Feminine(body string) Result { return checkFeminine(body) }
+
+// HisGender — see checkHisGender.
+func HisGender(body string) Result { return checkHisGender(body) }
+
+// Address — see checkAddress.
+func Address(body string) Result { return checkAddress(body) }
+
 func checkMood(mood string) Result {
 	if Moods[mood] {
 		return Result{CheckMood, true, ""}