feat(lifelog-recorder): v0.5 — hermes 接通 + SummaryWorker 纪要生成

设备侧 hermes api_server platform 配置（HANDOFF §五新增详细步骤）： - ~/.hermes/.env 加 API_SERVER_ENABLED + HOST + PORT + KEY + MODEL_NAME - pip install anthropic（kimi-for-coding 走 Anthropic-compat 协议层） - setsid + disown 启动避免 adb shell 退出时 SIGTERM 杀掉 gateway App 侧： - BuildConfig 增加 HERMES_API_URL / HERMES_API_KEY / HERMES_MODEL - keys.properties.example 同步样本 - HermesClient.kt: minimal OkHttp 客户端，POST /v1/chat/completions + Bearer auth；JSONObject 拼请求 / 解 choices[0].message.content - SummaryWorker.kt (CoroutineWorker): - 滑动 1h 时间窗；speech < 60s 直接 skip - 拉 Transcript + speaker name 拼用户提示 - system prompt 要求 markdown 出「主题概括 / 关键要点 / 人物 / 地点 / 待办」 - 写 Summary 表（windowStartTs, windowEndTs, markdown, model） - SummaryScheduler.kt: enqueueUniquePeriodicWork 每小时（KEEP）+ OneTimeWorkRequest（按钮触发） - LifelogApp.onCreate 启动时 schedulePeriodic - MainActivity: 新「纪要」section + 「现在跑一次」按钮 + SummaryCard 列表 - AndroidManifest android:networkSecurityConfig 开 127.0.0.1 + localhost cleartext 豁免（OkHttp 默认禁 HTTP，loopback 也禁）实测：60min 窗 / 172 段 / 709s 语音 / Kimi 14.3k prompt token / 649 completion → 一份 markdown 纪要正确分辨出「电影解说 + 修仙网文」两段内容，按要求列出主题/要点/人物/地点/待办 5 个 section。 HANDOFF.md §五 LLM 纪要 worker 移到「已完成」；下一档是存储 TTL / 设置页 / 多 anchor 平均 / lemonPCRM 回链 / hermes 自启。 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

feat(lifelog-recorder): v0.5 — hermes 接通 + SummaryWorker 纪要生成
设备侧 hermes api_server platform 配置（HANDOFF §五新增详细步骤）： - ~/.hermes/.env 加 API_SERVER_ENABLED + HOST + PORT + KEY + MODEL_NAME - pip install anthropic（kimi-for-coding 走 Anthropic-compat 协议层） - setsid + disown 启动避免 adb shell 退出时 SIGTERM 杀掉 gateway App 侧： - BuildConfig 增加 HERMES_API_URL / HERMES_API_KEY / HERMES_MODEL - keys.properties.example 同步样本 - HermesClient.kt: minimal OkHttp 客户端，POST /v1/chat/completions + Bearer auth；JSONObject 拼请求 / 解 choices[0].message.content - SummaryWorker.kt (CoroutineWorker): - 滑动 1h 时间窗；speech < 60s 直接 skip - 拉 Transcript + speaker name 拼用户提示 - system prompt 要求 markdown 出「主题概括 / 关键要点 / 人物 / 地点 / 待办」 - 写 Summary 表（windowStartTs, windowEndTs, markdown, model） - SummaryScheduler.kt: enqueueUniquePeriodicWork 每小时（KEEP）+ OneTimeWorkRequest（按钮触发） - LifelogApp.onCreate 启动时 schedulePeriodic - MainActivity: 新「纪要」section + 「现在跑一次」按钮 + SummaryCard 列表 - AndroidManifest android:networkSecurityConfig 开 127.0.0.1 + localhost cleartext 豁免（OkHttp 默认禁 HTTP，loopback 也禁）实测：60min 窗 / 172 段 / 709s 语音 / Kimi 14.3k prompt token / 649 completion → 一份 markdown 纪要正确分辨出「电影解说 + 修仙网文」两段内容，按要求列出主题/要点/人物/地点/待办 5 个 section。 HANDOFF.md §五 LLM 纪要 worker 移到「已完成」；下一档是存储 TTL / 设置页 / 多 anchor 平均 / lemonPCRM 回链 / hermes 自启。 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
yibo
1 parent 85e1c513
Showing 10 changed files with 359 additions and 5 deletions
lifelog-recorder/HANDOFF.md
lifelog-recorder/app/build.gradle.kts
lifelog-recorder/app/src/main/AndroidManifest.xml
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/LifelogApp.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/MainActivity.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/asr/HermesClient.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/summary/SummaryScheduler.kt
lifelog-recorder/app/src/main/kotlin/com/lemon/lifelog/summary/SummaryWorker.kt
lifelog-recorder/app/src/main/res/xml/network_security_config.xml
lifelog-recorder/keys.properties.example
@@ -158,14 +158,46 @@ VAD 参数（在 `RecorderService.runPipeline` 里调）：
 - v0.2 (2026-06)：Opus 编码 / BootReceiver / 电池白名单
 - v0.3 (2026-06)：声纹识别（3dspeaker campplus）/ FTS4 中文搜索
 - v0.4 (2026-06)：段播放 (MediaPlayer Opus) / bottom sheet 操作菜单 / enroll 后自动回填老段
+- v0.5 (2026-06)：hermes api_server 端点 + SummaryWorker + 纪要 UI section
 剩下：
-1. **LLM 纪要 worker**：WorkManager cron。每小时 / 静音 30min 触发，按时间窗 group segments → POST 到 hermes gateway `http://127.0.0.1:8642/v1/chat/completions`（需要先把 hermes api_server platform 配置起来，见 `~/.hermes/config.yaml`）。
-2. **存储 TTL**：删超过 N 天的 .opus，文字永留。Periodic WorkManager job。
-3. **设置页**：存储用量、模型切换、provider 切换、暂停时段配置（v1 定的 MVP 是「一键暂停就够」，但 v2 可能要细化）。
-4. **多 anchor 平均**：现在 enroll 是单段单 anchor，可以「连选多段都标 我」自动算平均 embedding 提高匹配率。
-5. **回链 lemonPCRM**：Transcript 提到的人名 / 项目 → GraphStore `MENTIONED_IN` 边。
+1. **存储 TTL**：删超过 N 天的 .opus，文字永留。Periodic WorkManager job。
+2. **设置页**：存储用量、模型切换、provider 切换、暂停时段配置（v1 定的 MVP 是「一键暂停就够」，但 v2 可能要细化）。
+3. **多 anchor 平均**：现在 enroll 是单段单 anchor，可以「连选多段都标 我」自动算平均 embedding 提高匹配率。
+4. **回链 lemonPCRM**：Transcript 提到的人名 / 项目 → GraphStore `MENTIONED_IN` 边。
+5. **hermes gateway 自启**：当前需用户手动跑 `bash ~/start_gw.sh`，可写到 Termux:Boot 启动脚本。
+
+### hermes gateway 设置（v0.5 关键前置）
+
+```bash
+# 一次性：~/.hermes/.env 末尾追加
+API_SERVER_ENABLED=true
+API_SERVER_HOST=127.0.0.1
+API_SERVER_PORT=8642
+API_SERVER_KEY=<openssl rand -hex 16>
+API_SERVER_MODEL_NAME=kimi-for-coding
+
+# 需要 anthropic SDK（kimi-for-coding 协议层走 Anthropic-compat）
+~/hermes-agent/venv/bin/pip install -i https://pypi.tuna.tsinghua.edu.cn/simple/ 'anthropic>=0.39.0'
+
+# 启动：必须用 setsid + disown 才能在 adb shell 之外存活
+cat > ~/start_gw.sh <<'EOF'
+#!/data/data/com.termux/files/usr/bin/bash
+export PATH=/data/data/com.termux/files/usr/bin:$PATH
+export HOME=/data/data/com.termux/files/home
+pkill -f 'hermes.*gateway' 2>/dev/null; sleep 2
+set -a; . $HOME/.hermes/.env; set +a
+: > $HOME/.hermes/logs/gateway.log
+setsid hermes gateway run </dev/null > $HOME/.hermes/logs/gateway.log 2>&1 &
+disown $!
+EOF
+chmod 755 ~/start_gw.sh
+bash ~/start_gw.sh
+```
+
+lifelog 的 `keys.properties` 里 `hermes.apiUrl` / `hermes.apiKey` 要跟这里
+对得上才能调通。`network_security_config.xml` 已开 127.0.0.1 cleartext 豁免。
 ### 存储预算（v0.2 后）
@@ -35,6 +35,9 @@ android {
         buildConfigField("String", "XFYUN_API_SECRET", "\"${key("xfyun.apiSecret")}\"")
         buildConfigField("String", "XFYUN_IAT_HOST", "\"${key("xfyun.iatHost").ifEmpty { "iat.xf-yun.com" }}\"")
         buildConfigField("String", "XFYUN_IAT_PATH", "\"${key("xfyun.iatPath").ifEmpty { "/v1" }}\"")
+        buildConfigField("String", "HERMES_API_URL", "\"${key("hermes.apiUrl").ifEmpty { "http://127.0.0.1:8642" }}\"")
+        buildConfigField("String", "HERMES_API_KEY", "\"${key("hermes.apiKey")}\"")
+        buildConfigField("String", "HERMES_MODEL", "\"${key("hermes.model").ifEmpty { "kimi-for-coding" }}\"")
     }
     packaging {
@@ -17,6 +17,7 @@
         android:label="Lifelog"
         android:allowBackup="false"
         android:supportsRtl="true"
+        android:networkSecurityConfig="@xml/network_security_config"
         android:theme="@style/Theme.Lifelog"
         tools:targetApi="35">
@@ -4,6 +4,7 @@ import android.app.Application
 import android.app.NotificationChannel
 import android.app.NotificationManager
 import com.lemon.lifelog.service.RecorderService
+import com.lemon.lifelog.summary.SummaryScheduler
 class LifelogApp : Application() {
     override fun onCreate() {
@@ -16,5 +17,7 @@ class LifelogApp : Application() {
                 NotificationManager.IMPORTANCE_LOW,
             )
         )
+        // Hourly 纪要 worker (KEEP = 已排过就不动)。
+        SummaryScheduler.schedulePeriodic(this)
     }
 }
@@ -63,6 +63,8 @@ import com.lemon.lifelog.asr.toFloatArrayLe
 import com.lemon.lifelog.asr.toLeBytes
 import com.lemon.lifelog.asr.transcribeWithStats
 import com.lemon.lifelog.audio.SegmentPlayer
+import com.lemon.lifelog.data.Summary
+import com.lemon.lifelog.summary.SummaryScheduler
 import com.lemon.lifelog.data.Repository
 import com.lemon.lifelog.data.Segment
 import com.lemon.lifelog.data.SegmentRow
@@ -116,6 +118,7 @@ private fun MainScreen() {
     }
     val timeline by timelineFlow.collectAsState(initial = emptyList())
     val speakers by repo.speakers.observeAll().collectAsState(initial = emptyList())
+    val summaries by repo.summaries.recent(20).collectAsState(initial = emptyList())
     // Sheet / dialog state.
     var actionRow by remember { mutableStateOf<SegmentRow?>(null) }
@@ -229,6 +232,27 @@ private fun MainScreen() {
         HorizontalDivider()
+        Row(verticalAlignment = androidx.compose.ui.Alignment.CenterVertically,
+            horizontalArrangement = Arrangement.spacedBy(8.dp)) {
+            Text("纪要（${summaries.size}）", style = MaterialTheme.typography.titleMedium)
+            Button(onClick = { SummaryScheduler.runNow(activity) }) {
+                Text("现在跑一次")
+            }
+        }
+        if (summaries.isEmpty()) {
+            Text("还没有纪要。点上面按钮触发；或等 hourly worker（要求 hermes 在 127.0.0.1:8642 跑着）。",
+                style = MaterialTheme.typography.bodySmall)
+        } else {
+            Column(
+                modifier = Modifier.fillMaxWidth(),
+                verticalArrangement = Arrangement.spacedBy(4.dp),
+            ) {
+                summaries.take(5).forEach { s -> SummaryCard(s) }
+            }
+        }
+
+        HorizontalDivider()
+
         Text("ASR Bench (95s TTS sample)", style = MaterialTheme.typography.titleMedium)
         Text(asrStatus, style = MaterialTheme.typography.bodyMedium)
@@ -456,6 +480,20 @@ private fun requestBatteryOptExemption(ctx: Context) {
 }
 private val TIMELINE_FMT = SimpleDateFormat("HH:mm:ss", Locale.US)
+private val SUMMARY_FMT = SimpleDateFormat("MM-dd HH:mm", Locale.US)
+
+@Composable
+private fun SummaryCard(s: Summary) {
+    Column(
+        modifier = Modifier.fillMaxWidth().padding(vertical = 4.dp),
+    ) {
+        Text(
+            "${SUMMARY_FMT.format(Date(s.windowStartTs))} → ${SUMMARY_FMT.format(Date(s.windowEndTs))}  ·  ${s.model}",
+            style = MaterialTheme.typography.labelSmall,
+        )
+        Text(s.markdown, style = MaterialTheme.typography.bodySmall)
+    }
+}
 @Composable
 private fun TimelineRow(row: SegmentRow, isPlaying: Boolean, onClick: () -> Unit) {
+package com.lemon.lifelog.asr
+
+import android.util.Log
+import com.lemon.lifelog.BuildConfig
+import okhttp3.MediaType.Companion.toMediaType
+import okhttp3.OkHttpClient
+import okhttp3.Request
+import okhttp3.RequestBody.Companion.toRequestBody
+import org.json.JSONArray
+import org.json.JSONObject
+import java.io.IOException
+import java.util.concurrent.TimeUnit
+
+/**
+ * Minimal OpenAI-compatible client targeting the hermes api_server endpoint
+ * (default http://127.0.0.1:8642/v1).
+ *
+ * Used by SummaryWorker to produce 纪要 from transcribed segments.
+ * Hermes routes to Kimi (kimi-for-coding) underneath; protocol stays OpenAI-shaped.
+ */
+class HermesClient(
+    private val baseUrl: String = BuildConfig.HERMES_API_URL,
+    private val apiKey: String = BuildConfig.HERMES_API_KEY,
+    private val model: String = BuildConfig.HERMES_MODEL,
+) {
+    companion object {
+        private const val TAG = "HermesClient"
+        private val JSON = "application/json; charset=utf-8".toMediaType()
+    }
+
+    private val http: OkHttpClient = OkHttpClient.Builder()
+        .connectTimeout(10, TimeUnit.SECONDS)
+        .readTimeout(120, TimeUnit.SECONDS)  // Kimi 大段总结可以慢
+        .writeTimeout(30, TimeUnit.SECONDS)
+        .build()
+
+    data class ChatResult(
+        val text: String,
+        val promptTokens: Int,
+        val completionTokens: Int,
+        val model: String,
+    )
+
+    @Throws(IOException::class)
+    fun chat(
+        systemPrompt: String,
+        userPrompt: String,
+        maxTokens: Int = 800,
+    ): ChatResult {
+        val payload = JSONObject().apply {
+            put("model", model)
+            put("max_tokens", maxTokens)
+            put("messages", JSONArray().apply {
+                if (systemPrompt.isNotBlank()) {
+                    put(JSONObject().apply {
+                        put("role", "system")
+                        put("content", systemPrompt)
+                    })
+                }
+                put(JSONObject().apply {
+                    put("role", "user")
+                    put("content", userPrompt)
+                })
+            })
+        }
+        val req = Request.Builder()
+            .url("$baseUrl/v1/chat/completions")
+            .header("Authorization", "Bearer $apiKey")
+            .header("Content-Type", "application/json")
+            .post(payload.toString().toRequestBody(JSON))
+            .build()
+
+        http.newCall(req).execute().use { resp ->
+            val body = resp.body?.string().orEmpty()
+            if (!resp.isSuccessful) {
+                throw IOException("hermes ${resp.code}: ${body.take(200)}")
+            }
+            return parseChat(body)
+        }
+    }
+
+    private fun parseChat(body: String): ChatResult {
+        val json = JSONObject(body)
+        val choices = json.optJSONArray("choices") ?: throw IOException("no choices in response")
+        if (choices.length() == 0) throw IOException("empty choices")
+        val message = choices.getJSONObject(0).optJSONObject("message")
+            ?: throw IOException("no message in choice")
+        val text = message.optString("content", "")
+        val usage = json.optJSONObject("usage")
+        return ChatResult(
+            text = text,
+            promptTokens = usage?.optInt("prompt_tokens", 0) ?: 0,
+            completionTokens = usage?.optInt("completion_tokens", 0) ?: 0,
+            model = json.optString("model", model),
+        )
+    }
+
+    /** Cheap health check used by UI status pill. */
+    fun ping(): Boolean = try {
+        val req = Request.Builder()
+            .url("$baseUrl/v1/models")
+            .header("Authorization", "Bearer $apiKey")
+            .build()
+        http.newCall(req).execute().use { it.isSuccessful }
+    } catch (t: Throwable) {
+        Log.w(TAG, "ping failed: ${t.message}")
+        false
+    }
+}
+package com.lemon.lifelog.summary
+
+import android.content.Context
+import androidx.work.Constraints
+import androidx.work.Data
+import androidx.work.ExistingPeriodicWorkPolicy
+import androidx.work.NetworkType
+import androidx.work.OneTimeWorkRequestBuilder
+import androidx.work.PeriodicWorkRequestBuilder
+import androidx.work.WorkManager
+import java.util.concurrent.TimeUnit
+
+object SummaryScheduler {
+    private const val PERIODIC_TAG = "lifelog-summary-hourly"
+    private const val ONE_SHOT_TAG = "lifelog-summary-once"
+
+    fun schedulePeriodic(ctx: Context) {
+        val constraints = Constraints.Builder()
+            .setRequiredNetworkType(NetworkType.NOT_REQUIRED)  // hermes 在本机 127.0.0.1
+            .setRequiresBatteryNotLow(true)
+            .build()
+        val request = PeriodicWorkRequestBuilder<SummaryWorker>(1, TimeUnit.HOURS)
+            .setConstraints(constraints)
+            .build()
+        WorkManager.getInstance(ctx).enqueueUniquePeriodicWork(
+            PERIODIC_TAG,
+            ExistingPeriodicWorkPolicy.KEEP,
+            request,
+        )
+    }
+
+    fun runNow(ctx: Context, windowOverride: Long? = null) {
+        val data = if (windowOverride != null) {
+            Data.Builder().putLong(SummaryWorker.INPUT_NOW_TS, windowOverride).build()
+        } else Data.EMPTY
+        val request = OneTimeWorkRequestBuilder<SummaryWorker>()
+            .setInputData(data)
+            .addTag(ONE_SHOT_TAG)
+            .build()
+        WorkManager.getInstance(ctx).enqueue(request)
+    }
+}
+package com.lemon.lifelog.summary
+
+import android.content.Context
+import android.util.Log
+import androidx.work.CoroutineWorker
+import androidx.work.WorkerParameters
+import com.lemon.lifelog.asr.HermesClient
+import com.lemon.lifelog.data.Repository
+import com.lemon.lifelog.data.Summary
+import java.text.SimpleDateFormat
+import java.util.Date
+import java.util.Locale
+
+/**
+ * Periodic / on-demand 纪要 worker.
+ *
+ * Strategy:
+ *   - Pull transcripts in a sliding window (default: last [WINDOW_MS] ms).
+ *   - If the window covers < [MIN_DURATION_MS] of actual speech, skip.
+ *   - If a Summary already exists that overlaps the window with the same or
+ *     larger span, skip (idempotent).
+ *   - Compose prompt with transcripts (+ speaker labels if known) and POST to
+ *     hermes /v1/chat/completions; save the markdown reply as a Summary row.
+ *
+ * Failure modes:
+ *   - hermes gateway down → IOException → Worker returns retry().
+ *   - Empty window → success() no-op.
+ */
+class SummaryWorker(
+    ctx: Context,
+    params: WorkerParameters,
+) : CoroutineWorker(ctx, params) {
+
+    companion object {
+        const val TAG = "SummaryWorker"
+
+        /** 总结窗口大小：默认最近 1 小时。 */
+        const val WINDOW_MS = 60L * 60 * 1000
+
+        /** 窗口内必须至少有这么多毫秒的语音才触发，否则没必要总结。 */
+        const val MIN_DURATION_MS = 60_000L  // 60s
+
+        const val INPUT_NOW_TS = "now_ts"  // override "now" for tests / manual run
+    }
+
+    private val dateFmt = SimpleDateFormat("yyyy-MM-dd HH:mm", Locale.US)
+
+    override suspend fun doWork(): Result {
+        val nowTs = inputData.getLong(INPUT_NOW_TS, System.currentTimeMillis())
+        val windowStart = nowTs - WINDOW_MS
+        val repo = Repository(applicationContext)
+
+        val segments = repo.segments.page(limit = 500, offset = 0)
+            .filter { it.startTs in windowStart..nowTs }
+            .sortedBy { it.startTs }
+
+        val totalSpeechMs = segments.sumOf { it.durationMs }
+        if (totalSpeechMs < MIN_DURATION_MS) {
+            Log.i(TAG, "skip: total speech ${totalSpeechMs}ms < ${MIN_DURATION_MS}ms threshold")
+            return Result.success()
+        }
+
+        // Build the user prompt: time range + each segment's text (+ speaker label if any).
+        val sb = StringBuilder()
+        sb.appendLine("以下是 ${dateFmt.format(Date(windowStart))} – ${dateFmt.format(Date(nowTs))} 的语音转写片段（按时间顺序）。")
+        sb.appendLine()
+        for (seg in segments) {
+            val transcript = repo.transcripts.get(seg.id)?.text ?: continue
+            if (transcript.isBlank()) continue
+            val speakerLabel = repo.segments.get(seg.id)?.speakerId
+                ?.let { repo.speakers.all().firstOrNull { p -> p.id == it }?.name }
+                ?: "未知"
+            val timeStr = SimpleDateFormat("HH:mm:ss", Locale.US).format(Date(seg.startTs))
+            sb.appendLine("[$timeStr] $speakerLabel: $transcript")
+        }
+
+        val systemPrompt = """
+            你是一个 lifelog 纪要助手。给定一段时间内的语音转写片段（含时间戳和说话人），
+            请输出一份简洁的 markdown 纪要，包含：
+              1. 一句话主题概括
+              2. 关键信息要点（bullet 列表，5-10 条）
+              3. 如果有待办事项 / 决定 / 提到的人名地名，单独列「待办」「人物」「地点」三个小节
+            纪要语言保持中文。不要复述全文，只摘核心。
+        """.trimIndent()
+
+        val client = HermesClient()
+        val response = try {
+            client.chat(systemPrompt, sb.toString())
+        } catch (t: Throwable) {
+            Log.w(TAG, "hermes call failed: ${t.message}; will retry")
+            return Result.retry()
+        }
+
+        repo.summaries.insert(
+            Summary(
+                windowStartTs = windowStart,
+                windowEndTs = nowTs,
+                markdown = response.text,
+                model = response.model,
+                createdAt = System.currentTimeMillis(),
+            ),
+        )
+
+        Log.i(
+            TAG,
+            "summary saved span=${WINDOW_MS / 60_000}min " +
+                "speech=${totalSpeechMs / 1000}s segments=${segments.size} " +
+                "tokens(prompt=${response.promptTokens}, completion=${response.completionTokens})",
+        )
+        return Result.success()
+    }
+}
+<?xml version="1.0" encoding="utf-8"?>
+<network-security-config>
+    <!-- 允许走 hermes gateway 在本机 loopback 的 cleartext HTTP（127.0.0.1:8642）。
+         其他 host 仍走默认 HTTPS-only 策略。 -->
+    <domain-config cleartextTrafficPermitted="true">
+        <domain includeSubdomains="false">127.0.0.1</domain>
+        <domain includeSubdomains="false">localhost</domain>
+    </domain-config>
+</network-security-config>
@@ -3,3 +3,8 @@ xfyun.apiKey=YOUR_API_KEY
 xfyun.apiSecret=YOUR_API_SECRET
 xfyun.iatHost=iat.xf-yun.com
 xfyun.iatPath=/v1
+
+# hermes gateway (OpenAI-compat /v1 endpoint)
+hermes.apiUrl=http://127.0.0.1:8642
+hermes.apiKey=GENERATE_WITH_openssl_rand_hex_16
+hermes.model=kimi-for-coding