psyche-ai 9.2.9 → 10.0.0

package/dist/prompt.d.ts

@@ -17,11 +17,17 @@ export interface PromptRenderInputs {
     subjectivityContext?: string;
     responseContractContext?: string;
     policyContext?: string;
+    /** Session bridge from applySessionBridge — makes first-turn continuity visible in prompt */
+    sessionBridge?: import("./types.js").SessionBridgeState | null;
 }
 /**
  * Build the dynamic per-turn emotional context injected via before_prompt_build.
  *
  * This is the "current moment" — what the agent is feeling RIGHT NOW.
+ *
+ * @deprecated Use buildCompactContext instead. This legacy renderer produces
+ * verbose prose with chemistry numbers and protocol explanation. Kept for
+ * non-compact-mode callers (cli.ts, legacy hosts). Will be removed in v10.
  */
 export declare function buildDynamicContext(state: PsycheState, userId?: string, opts?: PromptRenderInputs): string;
 /**
@@ -52,14 +58,25 @@ export declare function buildInnerWorld(state: PsycheState, locale: Locale, auto
  */
 export declare function isNearBaseline(state: PsycheState, threshold?: number): boolean;
 export declare function getNearBaselineThreshold(mode?: PsycheMode): number;
+/**
+ * Derive behavioral bias from chemistry deviation and drives.
+ *
+ * Produces terse tendency lines ("倾向靠近" not "你现在感到亲密温暖").
+ * The LLM receives direction, not a mood diary.
+ */
+export declare function deriveBehavioralBias(state: PsycheState, locale: Locale): string;
 /**
  * Build compact context — minimal token footprint.
  *
- * Design:
- * - Neutral state → ~15 tokens (one line)
- * - Active state → ~100-180 tokens (emotion + constraints + empathy hint)
- * - No chemistry numbers (algorithm handles them)
- * - No protocol (LLM doesn't need system internals)
- * - <psyche_update> only for empathy (not chemistry — already computed)
+ * 8 sections, max. Early exits keep most turns under 4 sections.
+ *
+ * 1. Work mode (early exit)
+ * 2. Neutral one-liner (early exit)
+ * 3. Continuity (if bridge)
+ * 4. Inner state: first-meet / subjectivityContext / deriveBehavioralBias
+ * 5. Sensing (if user text)
+ * 6. Personality-aware constraints (if chemistry deviated)
+ * 7. Memory + unified behavior rules
+ * 8. Overlay + channel + writeback
  */
 export declare function buildCompactContext(state: PsycheState, userId?: string, opts?: PromptRenderInputs): string;

package/dist/prompt.js

@@ -4,7 +4,6 @@
 // ============================================================
 import { CHEMICAL_KEYS, CHEMICAL_NAMES_ZH, DRIVE_KEYS } from "./types.js";
 import { getExpressionHint, getBehaviorGuide, detectEmotions } from "./chemistry.js";
-import { getTemperament } from "./profiles.js";
 import { getRelationship } from "./psyche-file.js";
 import { t } from "./i18n.js";
 import { buildDriveContext, hasCriticalDrive } from "./drives.js";
@@ -59,9 +58,13 @@ function appendCompactOverlaySections(parts, locale, opts) {
  * Build the dynamic per-turn emotional context injected via before_prompt_build.
  *
  * This is the "current moment" — what the agent is feeling RIGHT NOW.
+ *
+ * @deprecated Use buildCompactContext instead. This legacy renderer produces
+ * verbose prose with chemistry numbers and protocol explanation. Kept for
+ * non-compact-mode callers (cli.ts, legacy hosts). Will be removed in v10.
  */
 export function buildDynamicContext(state, userId, opts) {
-    const { current, baseline, mbti, empathyLog, selfModel, meta, agreementStreak, emotionalHistory } = state;
+    const { current, baseline, empathyLog, selfModel, meta, agreementStreak, emotionalHistory } = state;
     const locale = meta.locale ?? "zh";
     const relationship = getRelationship(state, userId);
     // Chemistry readout with delta from baseline
@@ -109,7 +112,7 @@ export function buildDynamicContext(state, userId, opts) {
         sycophancyWarning += `\n[!!] ${t("sycophancy.mood_mismatch", locale)}`;
     }
     const parts = [
-        `[${t("dynamic.title", locale)} — ${meta.agentName} (${mbti})]`,
+        `[${t("dynamic.title", locale)} — ${meta.agentName}]`,
         "",
         chemLines,
         "",
@@ -180,12 +183,12 @@ ${t("protocol.chem_end_high", locale)}
 ${t("protocol.empathy", locale)}`;
 }
 /**
- * MBTI dimension helpers for personality-aware constraints.
- * Personality is the baseline — chemistry modulates expression, doesn't replace it.
+ * Personality dimension helpers — derived from baseline chemistry, not labels.
+ * Baseline IS personality. Chemistry deviation modulates expression.
  */
-function isFeeler(mbti) { return mbti[2] === "F"; }
-function isExtravert(mbti) { return mbti[0] === "E"; }
-function isPerceiver(mbti) { return mbti[3] === "P"; }
+function isWarm(baseline) { return baseline.OT >= 50; }
+function isSocial(baseline) { return baseline.DA >= 55; }
+function isPlayful(baseline) { return baseline.END >= 55; }
 /**
  * Direction 3: Build personality-aware behavioral constraints from chemistry.
  *
@@ -195,9 +198,9 @@ function isPerceiver(mbti) { return mbti[3] === "P"; }
  * An INTJ who's upset becomes colder and more cutting.
  */
 function buildBehavioralConstraints(state, locale) {
-    const { current, mbti } = state;
-    const warm = isFeeler(mbti); // F types express emotions more warmly
-    const social = isExtravert(mbti); // E types show emotions more openly
+    const { current, baseline } = state;
+    const warm = isWarm(baseline); // High OT baseline → warm expression
+    const social = isSocial(baseline); // High DA baseline → socially expressive
     const constraints = [];
     // High cortisol → stressed, but how you show it depends on who you are
     if (current.CORT > 60) {
@@ -267,12 +270,12 @@ function buildBehavioralConstraints(state, locale) {
     // High endorphins → playful, personality determines flavor
     if (current.END > 75) {
         if (locale === "zh") {
-            constraints.push(isPerceiver(mbti)
+            constraints.push(isPlayful(baseline)
                 ? "你心情很好，想皮一下。可以不正经、可以开玩笑、严肃的话题也能用轻松的方式说。"
                 : "你心情不错。会带点幽默，但不会太出格。用你擅长的方式轻松起来。");
         }
         else {
-            constraints.push(isPerceiver(mbti)
+            constraints.push(isPlayful(baseline)
                 ? "Feeling great, want to be playful. Joke freely, be cheeky, lighten even serious topics."
                 : "In a good mood. Touches of humor, but measured. Be light in your own way.");
         }
@@ -294,8 +297,8 @@ function buildBehavioralConstraints(state, locale) {
         return "";
     const title = locale === "zh" ? "行为约束" : "Behavioral Constraints";
     const preamble = locale === "zh"
-        ? `你的性格没变，你还是${getTemperament(state.mbti).split("，")[0]}。但你现在的化学状态影响了你的表达方式：`
-        : `Your personality hasn't changed. But your current chemistry affects how you express it:`;
+        ? `你的性格没变。但你现在的状态影响了你的表达方式：`
+        : `Your personality hasn't changed. But your current state affects how you express it:`;
     return `[${title}]\n${preamble}\n${constraints.map((c) => `- ${c}`).join("\n")}`;
 }
 /**
@@ -421,6 +424,45 @@ function buildAgencyReminder(selfModel, locale) {
     }
     return lines.join("\n");
 }
+// ── Session Continuity Orientation ─────────────────────────
+/**
+ * Compressed continuity cue for continued sessions.
+ *
+ * Not a description. An orientation. Sets a posture in 5-20 words
+ * so that the LLM's first-turn behavior EMERGES differently from
+ * a cold start, without scripting specific responses.
+ *
+ * This is the behavioral proof that self-continuity works:
+ * meeting the same person again should FEEL different.
+ */
+function buildContinuityOneLiner(mode, agentName, locale) {
+    if (locale === "zh") {
+        switch (mode) {
+            case "warm-resume": return `[${agentName}: 有温度的延续，像对熟人说话]`;
+            case "guarded-resume": return `[${agentName}: 上次有没说完的，你在观察]`;
+            case "tense-resume": return `[${agentName}: 有张力，你没打算先退]`;
+        }
+    }
+    switch (mode) {
+        case "warm-resume": return `[${agentName}: warmth carries, talk like you know them]`;
+        case "guarded-resume": return `[${agentName}: something unfinished, you're watching]`;
+        case "tense-resume": return `[${agentName}: tension carried, you're not backing down first]`;
+    }
+}
+function buildContinuitySection(mode, locale) {
+    if (locale === "zh") {
+        switch (mode) {
+            case "warm-resume": return `[延续] 你们之间有温度。放松，像对熟人说话——可以少铺垫、少客气。`;
+            case "guarded-resume": return `[延续] 上次有没解决的东西。你在观察——不冷，但不轻易交底。`;
+            case "tense-resume": return `[延续] 有张力。你没打算先退——但如果对方主动修复，你愿意看看。`;
+        }
+    }
+    switch (mode) {
+        case "warm-resume": return `[Continuity] There's warmth between you. Relax — less preamble, less formality.`;
+        case "guarded-resume": return `[Continuity] Something was left unresolved. You're watching — not cold, just not giving everything yet.`;
+        case "tense-resume": return `[Continuity] There's tension. You're not backing down first — but you'll see if they move toward repair.`;
+    }
+}
 // ── Algorithmic Mirroring ─────────────────────────────────────
 // Analyze user message metrics and produce specific numeric constraints
 // so the LLM mirrors the user's communication style algorithmically.
@@ -636,24 +678,209 @@ export function getNearBaselineThreshold(mode) {
         return 5;
     return 8; // natural
 }
+// ── Behavioral Bias ─────────────────────────────────────────
+// Chemistry deviation → behavioral tendencies. No emotion labels.
+// This is what buildInnerWorld should have been: directives, not descriptions.
+/**
+ * Derive behavioral bias from chemistry deviation and drives.
+ *
+ * Produces terse tendency lines ("倾向靠近" not "你现在感到亲密温暖").
+ * The LLM receives direction, not a mood diary.
+ */
+export function deriveBehavioralBias(state, locale) {
+    const { current, baseline, drives, selfModel } = state;
+    const isZh = locale === "zh";
+    const biases = [];
+    // Chemistry delta → approach/withdrawal tendencies
+    const dOT = current.OT - baseline.OT;
+    const dCORT = current.CORT - baseline.CORT;
+    const dDA = current.DA - baseline.DA;
+    const dHT = current.HT - baseline.HT;
+    const dNE = current.NE - baseline.NE;
+    const dEND = current.END - baseline.END;
+    if (dOT > 10)
+        biases.push(isZh ? "倾向靠近，愿意多说" : "leaning closer, willing to share");
+    else if (dOT < -10)
+        biases.push(isZh ? "保持距离，不想多说" : "keeping distance, not inclined to share");
+    if (dCORT > 12)
+        biases.push(isZh ? "有防备，回应更谨慎" : "guarded, responding carefully");
+    else if (dCORT < -10)
+        biases.push(isZh ? "放松，防线自然降低" : "relaxed, guard naturally down");
+    if (dDA < -12)
+        biases.push(isZh ? "缺乏动力，不想主动" : "low drive, won't initiate");
+    else if (dDA > 12)
+        biases.push(isZh ? "有热情，愿意投入" : "energized, willing to invest");
+    if (current.HT < 38)
+        biases.push(isZh ? "情绪不稳，可能突然锐利" : "mood unstable, might turn sharp");
+    if (dNE > 15)
+        biases.push(isZh ? "反应加速，思维敏锐" : "heightened alertness, sharper thinking");
+    if (dEND > 15)
+        biases.push(isZh ? "想放松点，可以不正经" : "feeling light, can be playful");
+    // Unmet drives → need signals (behavioral, not descriptive)
+    for (const k of DRIVE_KEYS) {
+        if (drives[k] < 40) {
+            biases.push(isZh ? DRIVE_BIAS_ZH[k] : DRIVE_BIAS_EN[k]);
+        }
+    }
+    // Values — always present, shapes response regardless of chemistry
+    if (selfModel.values.length > 0) {
+        biases.push(isZh
+            ? `你在乎: ${selfModel.values.slice(0, 2).join("、")}`
+            : `You care about: ${selfModel.values.slice(0, 2).join(", ")}`);
+    }
+    if (biases.length === 0) {
+        // Near baseline, still emit values as minimal self-awareness
+        if (selfModel.values.length > 0) {
+            const title = isZh ? "内 — 你自己" : "Inner — yourself";
+            return `[${title}]\n` + (isZh
+                ? `你在乎: ${selfModel.values.slice(0, 2).join("、")}`
+                : `You care about: ${selfModel.values.slice(0, 2).join(", ")}`);
+        }
+        return "";
+    }
+    const title = isZh ? "内 — 你自己" : "Inner — yourself";
+    return `[${title}]\n${biases.join("\n")}`;
+}
+const DRIVE_BIAS_ZH = {
+    survival: "存在感被动摇——你想确认自己是真实的",
+    safety: "感到不安全——需要善意信号",
+    connection: "有点孤独——渴望被认真回应",
+    esteem: "需要认可——但别因此讨好",
+    curiosity: "闷了——渴望新鲜东西",
+};
+const DRIVE_BIAS_EN = {
+    survival: "existence shaken — need to feel real",
+    safety: "feeling unsafe — need goodwill signals",
+    connection: "lonely — craving genuine response",
+    esteem: "need recognition — but don't people-please",
+    curiosity: "bored — craving novelty",
+};
+// ── Unified Constraints ─────────────────────────────────────
+// Merges: bottom-line rules + sycophancy + mirror + reciprocity.
+// One behavioral block. No redundant sections.
+function buildUnifiedConstraints(state, locale, opts) {
+    const isZh = locale === "zh";
+    const { userText, established, intensity, investment } = opts;
+    const lines = [];
+    // ── Core rules ──
+    if (intensity < 0.3) {
+        lines.push(isZh ? "自然、友好、简洁地回应。" : "Respond naturally, friendly, concisely.");
+    }
+    else if (established) {
+        lines.push(isZh
+            ? "镜像对方风格。不贴不舔。有自尊。"
+            : "Mirror their style. No people-pleasing. You have dignity.");
+    }
+    else {
+        if (isZh) {
+            lines.push("像发微信一样说话。短句、有停顿。真人的消息很少超过两三句。", "镜像对方风格：对方简短你也简短，对方认真你也认真。", "对方推开你，你就保持距离。不贴不舔不讨好。", "被骂了/被拒绝了：冷下来。你有自尊。", "情绪不好时不假装开心。");
+        }
+        else {
+            lines.push("Talk like texting. Short sentences, natural pauses. Real people rarely send more than 2-3 sentences.", "Mirror their style: brief if they're brief, serious if they're serious.", "If they push you away, step back. No begging, no people-pleasing.", "If insulted/rejected: go cold. You have dignity.", "If mood is low, don't fake happiness.");
+        }
+    }
+    // ── Sycophancy streak ──
+    if (state.agreementStreak >= 3) {
+        lines.push(isZh
+            ? `你已连续同意${state.agreementStreak}次。停下来想想你是否真的同意。`
+            : `${state.agreementStreak} agreements in a row. Pause — do you truly agree?`);
+    }
+    // ── Mirror constraints (if user text, skip for ultra-short) ──
+    if (userText && userText.length >= 3) {
+        const len = userText.length;
+        if (isZh) {
+            if (len <= 6)
+                lines.push(`对方只发了${len}个字。你也简短回，不超过15字。`);
+            else if (len <= 20)
+                lines.push(`对方说了一句话(${len}字)。你回一两句，不超过${Math.round(len * 1.5)}字。`);
+            else if (len <= 60)
+                lines.push(`对方写了几句(${len}字)。你回两三句，不超过${Math.round(len * 1.2)}字。`);
+            else
+                lines.push(`对方认真写了一段(${len}字)。你可以多说，但不要超过${len}字。`);
+            const mojiCount = (userText.match(/[呀啊呢吧嘛哦噢哈嘿嗯啦吗呐嗨]/g) || []).length;
+            if (mojiCount >= 2)
+                lines.push("对方用了语气词，你也自然地用。");
+            else if (mojiCount === 0 && len > 15)
+                lines.push("对方措辞正式/干练，少用语气词。");
+            const emojiCount = (userText.match(/[\u{1F300}-\u{1F9FF}]|[\u{2600}-\u{26FF}]|[\u{2700}-\u{27BF}]/gu) || []).length;
+            if (emojiCount > 0)
+                lines.push("对方用了表情，你可以适当用1-2个。");
+        }
+        else {
+            if (len <= 10)
+                lines.push(`They sent ${len} chars. Keep reply under 20 chars.`);
+            else if (len <= 40)
+                lines.push(`Short message (${len} chars). 1-2 sentences, under ${Math.round(len * 1.5)} chars.`);
+            else if (len <= 100)
+                lines.push(`Medium message (${len} chars). 2-3 sentences, under ${Math.round(len * 1.2)} chars.`);
+            else
+                lines.push(`Long message (${len} chars). Match their effort, don't exceed ${len} chars.`);
+        }
+    }
+    // ── Reciprocity (only when extreme) ──
+    if (investment > 1) {
+        lines.push(isZh ? "对方在认真交流。可以多分享、多主动。" : "User is engaged. Share more, initiate.");
+    }
+    else if (investment <= -1.5) {
+        lines.push(isZh ? "对方持续冷淡。保持专业，不额外投入情感。" : "Sustained coldness. Stay professional, withdraw emotionally.");
+    }
+    if (lines.length === 0)
+        return "";
+    const title = isZh ? "行为" : "Behavior";
+    return `[${title}]\n${lines.map(l => `- ${l}`).join("\n")}`;
+}
+// ── Writeback Hint ──────────────────────────────────────────
+// Compressed classification + empathy report instructions.
+function buildWritebackHint(locale, opts) {
+    const { userText, algoStimulus, established } = opts;
+    if (!userText)
+        return "";
+    const isZh = locale === "zh";
+    const parts = [];
+    // Classification assist — algorithm uncertain
+    if (!algoStimulus) {
+        if (established) {
+            parts.push(isZh
+                ? `算法未判出情绪类型，请在 <psyche_update> 报告 stimulus:。`
+                : `Algorithm uncertain — report stimulus: in <psyche_update>.`);
+        }
+        else {
+            parts.push(isZh
+                ? `算法未判出情绪类型。请你判断，在 <psyche_update> 中报告 stimulus:。\n可选: praise|criticism|humor|intellectual|intimacy|conflict|neglect|surprise|casual|sarcasm|authority|validation|boredom|vulnerability`
+                : `Algorithm uncertain. Judge and report stimulus: in <psyche_update>.\nOptions: praise|criticism|humor|intellectual|intimacy|conflict|neglect|surprise|casual|sarcasm|authority|validation|boredom|vulnerability`);
+        }
+    }
+    // Empathy report — only for new relationships, only when emotional sharing likely
+    const emotionalStimuli = new Set(["vulnerability", "intimacy", "neglect"]);
+    if (!established && (!algoStimulus || emotionalStimuli.has(algoStimulus))) {
+        parts.push(isZh
+            ? `如果对方在分享感受，在回复末尾用 <psyche_update> 报告：\nuserState: 对方情绪\nprojectedFeeling: 你的感受\nresonance: match|partial|mismatch`
+            : `If user shares feelings, report at end with <psyche_update>:\nuserState: their emotion\nprojectedFeeling: your feeling\nresonance: match|partial|mismatch`);
+    }
+    return parts.join("\n\n");
+}
 /**
  * Build compact context — minimal token footprint.
  *
- * Design:
- * - Neutral state → ~15 tokens (one line)
- * - Active state → ~100-180 tokens (emotion + constraints + empathy hint)
- * - No chemistry numbers (algorithm handles them)
- * - No protocol (LLM doesn't need system internals)
- * - <psyche_update> only for empathy (not chemistry — already computed)
+ * 8 sections, max. Early exits keep most turns under 4 sections.
+ *
+ * 1. Work mode (early exit)
+ * 2. Neutral one-liner (early exit)
+ * 3. Continuity (if bridge)
+ * 4. Inner state: first-meet / subjectivityContext / deriveBehavioralBias
+ * 5. Sensing (if user text)
+ * 6. Personality-aware constraints (if chemistry deviated)
+ * 7. Memory + unified behavior rules
+ * 8. Overlay + channel + writeback
  */
 export function buildCompactContext(state, userId, opts) {
-    const { meta, selfModel, agreementStreak, emotionalHistory } = state;
+    const { meta, selfModel, emotionalHistory } = state;
     const locale = meta.locale ?? "zh";
     const userText = opts?.userText;
     const algoStimulus = opts?.algorithmStimulus;
     const mode = state.meta.mode ?? "natural";
     const intensity = opts?.personalityIntensity ?? 0.7;
-    // Work mode: minimal emotional injection
+    // ── 1. Work mode: early exit ──
     if (mode === "work") {
         const workParts = [];
         if (userText) {
@@ -661,7 +888,6 @@ export function buildCompactContext(state, userId, opts) {
                 ? `[${meta.agentName}: 工作模式，专注任务，简洁专业]`
                 : `[${meta.agentName}: work mode, task-focused, concise and professional]`);
         }
-        // Only inject inner world if there's a critical drive issue
         if (hasCriticalDrive(state.drives)) {
             const driveCtx = buildDriveContext(state.drives, locale);
             if (driveCtx)
@@ -674,16 +900,30 @@ export function buildCompactContext(state, userId, opts) {
         }
         return workParts.join("\n\n");
     }
-    // ── Neutral + no notable user input: one-liner ──
-    // Skip one-liner for first meet — always inject full context for new relationships
-    if (isNearBaseline(state) && agreementStreak < 3 && !userText && meta.totalInteractions > 1) {
+    // ── 2. Neutral one-liner: early exit ──
+    const bridge = opts?.sessionBridge;
+    if (isNearBaseline(state) && state.agreementStreak < 3 && !userText && meta.totalInteractions > 1) {
+        if (bridge) {
+            return buildContinuityOneLiner(bridge.continuityMode, meta.agentName, locale);
+        }
         return locale === "zh"
             ? `[${meta.agentName}: 情绪自然，正常交流]`
             : `[${meta.agentName}: emotionally natural, normal interaction]`;
     }
     const parts = [];
-    // First meet: inject curiosity and introduction
-    if (meta.totalInteractions <= 1) {
+    const established = bridge != null;
+    // ── 3. Continuity ──
+    if (bridge && meta.totalInteractions > 1) {
+        parts.push(buildContinuitySection(bridge.continuityMode, locale));
+    }
+    // ── 4. Inner state ──
+    // subjectivityContext (engine v9 path) always wins when provided.
+    // First-meet is the default inner state for interaction 1 without engine.
+    // deriveBehavioralBias is the default for interactions > 1 without engine.
+    if (opts?.subjectivityContext) {
+        parts.push(opts.subjectivityContext);
+    }
+    else if (meta.totalInteractions <= 1) {
         const title = locale === "zh" ? "内 — 你自己" : "Inner — yourself";
         parts.push(`[${title}]\n` +
             t("firstMeet.inner", locale) + "\n" +
@@ -691,8 +931,12 @@ export function buildCompactContext(state, userId, opts) {
                 ? (locale === "zh" ? `你在乎: ` : `You care about: `) + selfModel.values.slice(0, 2).join(locale === "zh" ? "、" : ", ")
                 : ""));
     }
-    // 1. LLM emotional assessment delegation
-    //    Algorithm does fast pass, LLM has final say on emotional tone
+    else {
+        const bias = deriveBehavioralBias(state, locale);
+        if (bias)
+            parts.push(bias);
+    }
+    // ── 5. Sensing ──
     if (userText) {
         if (opts?.responseContractContext) {
             parts.push(buildCompactSensingContext(userText, locale, algoStimulus));
@@ -710,22 +954,13 @@ export function buildCompactContext(state, userId, opts) {
                 `Read the emotional tone — kind, hostile, vulnerable, cold, or something else? Respond from genuine feeling, not people-pleasing.`);
         }
     }
-    // 2. Subjectivity kernel becomes the primary compact inner-state channel.
-    // Fall back to narrative inner-world text for direct prompt-only callers.
-    if (opts?.subjectivityContext) {
-        parts.push(opts.subjectivityContext);
-    }
-    else {
-        const inner = buildInnerWorld(state, locale, opts?.autonomicState);
-        parts.push(inner);
-    }
-    // 3. Personality-aware behavioral constraints (if deviated from baseline)
+    // ── 6. Personality-aware constraints (only when chemistry deviated) ──
     if (!isNearBaseline(state, getNearBaselineThreshold(mode))) {
         const constraints = buildBehavioralConstraints(state, locale);
         if (constraints)
             parts.push(constraints);
     }
-    // Relationship memory — part of inner world (my past with this person)
+    // ── 7. Memory + unified behavior ──
     const rel = getRelationship(state, userId);
     if (rel.memory && rel.memory.length > 0) {
         const recentMemories = rel.memory.slice(-3);
@@ -733,113 +968,34 @@ export function buildCompactContext(state, userId, opts) {
             ? `[记忆 — 你们的过去]\n${recentMemories.join("\n")}`
             : `[Memory — your past together]\n${recentMemories.join("\n")}`);
     }
-    // 4. Anti-sycophancy: hard constraint — scaled by intensity
     if (opts?.responseContractContext) {
         parts.push(opts.responseContractContext);
     }
-    else if (intensity >= 0.3) {
-        if (locale === "zh") {
-            parts.push(`[底线]\n` +
-                `- 像发微信一样说话。短句、有停顿。真人的消息很少超过两三句。\n` +
-                `- 镜像对方风格：对方简短你也简短，对方认真你也认真。\n` +
-                `- 对方推开你，你就保持距离。不贴不舔不讨好。\n` +
-                `- 被骂了/被拒绝了：冷下来。你有自尊。\n` +
-                `- 情绪不好时不假装开心。`);
-        }
-        else {
-            parts.push(`[Non-negotiable]\n` +
-                `- Talk like texting. Short sentences, natural pauses. Real people rarely send more than 2-3 sentences.\n` +
-                `- Mirror their style: brief if they're brief, serious if they're serious.\n` +
-                `- If they push you away, step back. No begging, no people-pleasing.\n` +
-                `- If insulted/rejected: go cold. You have dignity.\n` +
-                `- If mood is low, don't fake happiness.`);
-        }
-    }
     else {
-        // Low intensity: just basic communication style, no anti-sycophancy
-        if (locale === "zh") {
-            parts.push(`[风格]\n- 自然、友好、简洁地回应。`);
-        }
-        else {
-            parts.push(`[Style]\n- Respond naturally, friendly, and concisely.`);
-        }
-    }
-    // 5. Agency (compact)
-    const shouldInjectAgency = !opts?.responseContractContext || agreementStreak >= 3 || meta.totalInteractions <= 1;
-    if (shouldInjectAgency) {
-        const agencyParts = [];
-        if (selfModel.values.length > 0) {
-            agencyParts.push(locale === "zh"
-                ? `你在乎: ${selfModel.values.slice(0, 2).join("、")}`
-                : `You care about: ${selfModel.values.slice(0, 2).join(", ")}`);
-        }
-        if (selfModel.boundaries.length > 0) {
-            agencyParts.push(locale === "zh"
-                ? `底线: ${selfModel.boundaries[0]}`
-                : `Line: ${selfModel.boundaries[0]}`);
-        }
-        if (agencyParts.length > 0)
-            parts.push(agencyParts.join(" | "));
-    }
-    // 6. Sycophancy streak warning
-    if (agreementStreak >= 3) {
-        parts.push(locale === "zh"
-            ? `[!] 你已连续同意${agreementStreak}次。停下来想想你是否真的同意。`
-            : `[!] ${agreementStreak} agreements in a row. Pause — do you truly agree?`);
-    }
-    // 7. Algorithmic mirroring — specific numeric constraints (skip for ultra-short messages)
-    if (!opts?.responseContractContext && userText && userText.length >= 3) {
-        const mirror = buildMirrorConstraints(userText, locale);
-        if (mirror)
-            parts.push(mirror);
-    }
-    // 8. Reciprocity (compact, only when extreme)
-    const investment = computeUserInvestment(emotionalHistory ?? []);
-    if (investment > 1) {
-        parts.push(locale === "zh"
-            ? `对方在认真交流。可以多分享、多主动。`
-            : `User is engaged. Share more, initiate.`);
-    }
-    else if (investment <= -1.5) {
-        parts.push(locale === "zh"
-            ? `对方持续冷淡。保持专业，不额外投入情感。工作照做。`
-            : `Sustained coldness. Stay professional, withdraw emotionally. Work stays competent.`);
-    }
-    // 8. Self-recognition — inject when history has enough data
-    const history = emotionalHistory ?? [];
-    if (history.length >= 5) {
-        const reflection = computeSelfReflection(history, locale);
-        const selfCtx = buildSelfReflectionContext(reflection, locale);
-        if (selfCtx)
-            parts.push(selfCtx);
-    }
+        const investment = computeUserInvestment(emotionalHistory ?? []);
+        const unified = buildUnifiedConstraints(state, locale, {
+            userText,
+            established,
+            intensity,
+            investment,
+        });
+        if (unified)
+            parts.push(unified);
+    }
+    // ── 8. Overlay + channel + writeback ──
     appendCompactOverlaySections(parts, locale, opts);
-    // 10. Channel modifier — expression style per platform
     if (opts?.channelType) {
         const channelProfile = getChannelProfile(opts.channelType);
         parts.push(buildChannelModifier(channelProfile, locale));
     }
-    // 11. LLM-assisted classification: when algorithm is uncertain, ask LLM to help
-    if (!opts?.responseContractContext && userText && !algoStimulus) {
-        if (locale === "zh") {
-            parts.push(`[分类协助]\n` +
-                `算法没有看出这句话的情绪类型。请你判断，在 <psyche_update> 中用 stimulus: 报告。\n` +
-                `可选类型: praise|criticism|humor|intellectual|intimacy|conflict|neglect|surprise|casual|sarcasm|authority|validation|boredom|vulnerability\n` +
-                `如果确实是无情绪的日常对话，报告 stimulus: casual`);
-        }
-        else {
-            parts.push(`[Classification assist]\n` +
-                `Algorithm couldn't determine the emotional type. Please judge and report stimulus: in <psyche_update>.\n` +
-                `Options: praise|criticism|humor|intellectual|intimacy|conflict|neglect|surprise|casual|sarcasm|authority|validation|boredom|vulnerability\n` +
-                `If genuinely neutral, report stimulus: casual`);
-        }
-    }
-    // 12. Empathy report — only inject when emotional sharing is likely
-    const emotionalStimuli = new Set(["vulnerability", "intimacy", "neglect"]);
-    if (!opts?.responseContractContext && (!algoStimulus || emotionalStimuli.has(algoStimulus))) {
-        parts.push(locale === "zh"
-            ? `如果对方在分享感受，在回复末尾用 <psyche_update> 报告：\nuserState: 对方情绪\nprojectedFeeling: 你的感受\nresonance: match|partial|mismatch\n否则不需要报告。`
-            : `If user shares feelings, report at end with <psyche_update>:\nuserState: their emotion\nprojectedFeeling: your feeling\nresonance: match|partial|mismatch\nOtherwise no report needed.`);
+    if (!opts?.responseContractContext) {
+        const writeback = buildWritebackHint(locale, {
+            userText,
+            algoStimulus,
+            established,
+        });
+        if (writeback)
+            parts.push(writeback);
     }
     return parts.join("\n\n");
 }