feat(scripts): add cost monitor with daily summary + threshold alerts

.gitignore

@@ -186,3 +186,5 @@ docs/*/*
 # Ignore all memories (keep folder marker only)
 memories/*
 !memories/.gitkeep
+memories/vault/*
+!memories/vault/.gitkeep

CLAUDE.md

@@ -181,6 +181,11 @@ All skills are under `.claude/skills/`. Any agent can use any skill when relevan
 ## Consensus Memory
 
 - `memories/consensus.md` - cross-cycle baton; must be updated before cycle end
+- `memories/vault/` - persistent vector memory index (`index.json`), auto-built by
+  `scripts/core/memory_vault.py` from consensus snapshots + `docs/` outputs. Every
+  cycle auto-loop injects the top semantically-relevant blocks into the prompt so
+  agents can recall historical decisions long collapsed out of consensus.md. It is
+  generated automatically; no manual upkeep needed.
 - `docs/<role>/` - agent outputs
 - `projects/` - all created projects
 

INDEX.md

@@ -57,7 +57,8 @@
 | 守护 | `scripts/wsl/uninstall-wsl-daemon.sh` | 卸载 WSL daemon |
 | 守护 | `scripts/wsl/wsl-daemon-status.sh` | 查询 WSL daemon 状态 |
 | 守护 | `scripts/macos/install-daemon.sh` | macOS launchd 安装/卸载 |
-| 核心 | `scripts/core/auto-loop.sh` | 主循环执行、熔断、日志、共识更新 |
+| 核心 | `scripts/core/auto-loop.sh` | 主循环执行、熔断、日志、共识更新、向量记忆召回/沉淀 |
+| 记忆 | `scripts/core/memory_vault.py` | 向量记忆库：`index`(沉淀共识+docs)、`search`(语义召回)、`status`、`clear`。纯 Python 零依赖，TF-IDF+余弦 |
 | 核心 | `scripts/core/monitor.sh` | 核心状态/日志输出 |
 | 核心 | `scripts/core/stop-loop.sh` | 核心停止/暂停/恢复控制 |
 

PROMPT.md

@@ -8,6 +8,10 @@
 
 当前共识已预加载在本 prompt 末尾。如果没有，读 `memories/consensus.md`。
 
+> **长程记忆（向量库）**：本轮 prompt 还会附带一段从 `memories/vault/` 语义检索出的"相关历史记忆"（`## Highly-relevant past memory`）。它补充了 consensus 里被折叠掉的历史上下文——包括更早的决策、docs 里的方案细节。当前共识 `.md` 是最新/最权威的，检索出的历史仅作为背景参考，两者冲突时以共识为准。
+
+每轮 cycle 结束后，脚本会把最新的 consensus 快照与 `docs/` 产出分块向量化并存入 `memories/vault/index.json`，供后续周期语义召回。这是自动进行的，agent 无需手动维护 vault。
+
 ### 2. 决策
 
 - 有明确 Next Action → 执行它

README-ZH.md

@@ -41,7 +41,7 @@ daemon (launchd / systemd --user, 崩溃自重启)
         └── sleep → 下一轮
 ```
 
-每个周期是一次独立的 CLI 调用。`memories/consensus.md` 是唯一的跨周期状态——类似接力赛传棒。
+每个周期是一次独立的 CLI 调用。`memories/consensus.md` 是跨周期运行状态的接力棒。另外系统会自动把 consensus 快照与 `docs/` 产出向量化沉淀到长期记忆库 `memories/vault/`，并在每个周期语义注入相关历史供长期召回（见第 2 层）。
 
 ## 你该看哪一节（按平台）
 
@@ -188,7 +188,8 @@ Auto-Company 并非简单调用 LLM API，而是一个高度解耦的 **多智
 
 ### 第 2 层：编排与状态控制层 (Orchestration & State Machine)
 *   **永续主循环 (The Auto-Loop)**：通过 `scripts/core/auto-loop.sh` 控制的执行循环，让 AI 摆脱“单次对话”，实现 24/7 永续运行。
-*   **轻量级状态机 (Consensus Memory)**：放弃复杂的向量数据库或内存管理，将跨周期的上下文压缩为一个 Markdown 文件：`memories/consensus.md`。每次循环开始前读取，结束前必须重写，作为整个系统的“接力棒”。
+*   **轻量级状态机 (Consensus Memory)**：将跨周期的运行上下文压缩为一个 Markdown 文件：`memories/consensus.md`。每次循环开始前读取，结束前必须重写，作为整个系统的“接力棒”。
+*   **持久化向量记忆 (Long-term Recall)**：在接力棒之上再加一层。`scripts/core/memory_vault.py` 把 consensus 快照与 `docs/` 产出向量化沉淀到 `memories/vault/index.json`（纯 Python 的 TF-IDF + 余弦相似度，零外部依赖、近零成本）。每个新周期自动语义检索最相关的历史知识块注入 prompt，使 consensus 中已被折叠的历史决策与细节依然可召回。存储后端可替换：改 `_embed_chunk()` 即可接上 ChromaDB 或模型 embedding。
 *   **高可用容错机制 (Resilience & Self-Healing)**：内置熔断器 (连续错误触发冷却)、限流退避 (429 报错自动休眠) 和沙箱重置 (未成功输出有效共识时自动回滚)。
 
 ### 第 1 层：基础设施与执行引擎层 (Execution Engine & Infrastructure)

README.md

@@ -41,7 +41,7 @@ daemon (launchd / systemd --user, auto-restart on crash)
         └── sleep -> next cycle
 ```
 
-Each cycle is an independent CLI call. `memories/consensus.md` is the only cross-cycle state.
+Each cycle is an independent CLI call. `memories/consensus.md` is the cross-cycle running-state baton. A supplemental persistent **vector memory** (`memories/vault/`) is auto-built from consensus snapshots + `docs/` and injected semantically each cycle for long-term recall (see Layer 2).
 
 ## Where To Start (By Platform)
 
@@ -187,7 +187,8 @@ Auto-Company is not a simple LLM API wrapper, but a highly decoupled **Multi-Age
 
 ### Layer 2: Orchestration & State Machine
 *   **The Auto-Loop**: The execution loop controlled by `scripts/core/auto-loop.sh` frees the AI from "single-turn conversations", enabling 24/7 continuous operation.
-*   **Lightweight State Machine (Consensus Memory)**: Forgoes complex vector databases or memory management, compressing cross-cycle context into a single Markdown file: `memories/consensus.md`. Read before every cycle and rewritten before it ends, acting as the system's "baton".
+*   **Lightweight State Machine (Consensus Memory)**: Compresses cross-cycle running context into a single Markdown file: `memories/consensus.md` (read before every cycle, rewritten before it ends — the "baton").
+*   **Persistent Vector Memory (Long-term Recall)**: Layered on top of the baton. `scripts/core/memory_vault.py` indexes consensus snapshots + `docs/` outputs into `memories/vault/index.json` (pure-Python TF-IDF + cosine, zero external deps, ~zero cost). Each new cycle auto-retrieves the top semantically-relevant historical blocks and injects them into the prompt, so past decisions and details that consensus.md has collapsed away remain recallable. Swappable backend: replace `_embed_chunk()` to plug in ChromaDB or a model embedding.
 *   **Resilience & Self-Healing**: Built-in circuit breakers (cooldown triggered by consecutive errors), rate-limit backoff (auto-sleep on 429 errors), and sandbox reset (auto-rollback if a valid consensus is not output).
 
 ### Layer 1: Execution Engine & Infrastructure

dashboard/app.js

@@ -23,6 +23,15 @@ const els = {
   logText: document.getElementById("logText"),
   rawText: document.getElementById("rawText"),
 
+  vaultSubtitle: document.getElementById("vaultSubtitle"),
+  vaultStats: document.getElementById("vaultStats"),
+  vaultChart: document.getElementById("vaultChart"),
+  vaultLatest: document.getElementById("vaultLatest"),
+  vaultResults: document.getElementById("vaultResults"),
+  vaultQuery: document.getElementById("vaultQuery"),
+  btnVaultSearch: document.getElementById("btnVaultSearch"),
+  btnVaultRefresh: document.getElementById("btnVaultRefresh"),
+
   btnRefresh: document.getElementById("btnRefresh"),
   btnStart: document.getElementById("btnStart"),
   btnStop: document.getElementById("btnStop"),
@@ -236,6 +245,130 @@ async function fetchStatus() {
   els.latency.textContent = `Roundtrip: ${elapsed}ms`;
 }
 
+function drawVaultChart(bySource) {
+  const canvas = els.vaultChart;
+  if (!canvas || !bySource) return;
+  const ctx = canvas.getContext("2d");
+  const { width: W, height: H } = canvas;
+  ctx.clearRect(0, 0, W, H);
+
+  const entries = Object.entries(bySource).sort((a, b) => b[1] - a[1]).slice(0, 8);
+  if (entries.length === 0) {
+    ctx.fillStyle = "#5b6b7a";
+    ctx.font = "14px 'Rajdhani', sans-serif";
+    ctx.fillText("no chunks indexed yet", 16, H / 2);
+    return;
+  }
+
+  const max = Math.max(...entries.map((e) => e[1]), 1);
+  const padL = 14, padT = 14;
+  const rowH = (H - padT * 2) / entries.length;
+  const barMax = W - padL - 16;
+
+  // subtle axes
+  ctx.strokeStyle = "rgba(255,255,255,0.06)";
+  ctx.lineWidth = 1;
+  ctx.beginPath();
+  ctx.moveTo(padL, padT - 4);
+  ctx.lineTo(padL, H - padT);
+  ctx.moveTo(padL, H - padT);
+  ctx.lineTo(W - 8, H - padT);
+  ctx.stroke();
+
+  entries.forEach(([src, count], i) => {
+    const y = padT + i * rowH + rowH / 2;
+    const len = count / max * barMax;
+    const grad = ctx.createLinearGradient(padL, 0, padL + len, 0);
+    grad.addColorStop(0, "#5ee7df");
+    grad.addColorStop(1, "#7c5cff");
+    ctx.fillStyle = grad;
+    const r = 3;
+    ctx.beginPath();
+    ctx.roundRect(padL, y - 7, Math.max(len, 2), 14, r);
+    ctx.fill();
+
+    // label
+    ctx.fillStyle = "rgba(220,230,240,0.85)";
+    ctx.font = "600 12px 'Rajdhani', sans-serif";
+    let label = src.replace(/\.md$/i, "");
+    if (label.length > 18) label = "…" + label.slice(-17);
+    ctx.fillText(label, padL + Math.max(len, 2) + 8, y + 4);
+    ctx.fillStyle = "#5ee7df";
+    ctx.font = "700 12px 'Rajdhani', sans-serif";
+    ctx.fillText(String(count), W - 10, y + 4);
+  });
+}
+
+function truncateMiddle(text, n = 150) {
+  const t = String(text || "").replace(/\s+/g, " ").trim();
+  if (t.length <= n) return t;
+  return t.slice(0, n) + "…";
+}
+
+function renderVault(data) {
+  if (!data.ok || !data.stats) {
+    els.vaultSubtitle.textContent = "Vault not initialized yet — will appear after the first cycle.";
+    els.vaultStats.innerHTML = "<p class='muted'>Waiting for memories/vault/index.json…</p>";
+    drawVaultChart(null);
+    els.vaultLatest.innerHTML = "";
+    return;
+  }
+
+  const s = data.stats;
+  const lastIdx = data.timestamp ? "" : "";
+  els.vaultSubtitle.textContent =
+    `Indexed ${s.lastIndexed || "recently"} · ${s.sizeMb} MB`;
+
+  const cols = [
+    ["Chunks", s.chunks],
+    ["Sources", s.sources],
+    ["Distinct Terms", s.distinctTerms],
+    ["Size", `${s.sizeMb} MB`],
+  ];
+  els.vaultStats.innerHTML = cols
+    .map(([k, v]) => `<div class="vault-stat"><dt>${k}</dt><dd>${v}</dd></div>`)
+    .join("");
+
+  drawVaultChart(data.bySource || {});
+
+  const latest = data.latest || [];
+  if (latest.length) {
+    els.vaultLatest.innerHTML = `<div class="vault-latest-title">Latest entries</div>` +
+      latest.map((c) =>
+        `<div class="vault-entry"><span class="tag">${escapeHtml(c.source)}</span>` +
+        `<span class="muted mono">${escapeHtml(truncateMiddle(c.text, 110))}</span></div>`
+      ).join("");
+  } else {
+    els.vaultLatest.innerHTML = "";
+  }
+}
+
+function renderVaultSearch(data) {
+  const hits = (data.search && data.search.hits) || [];
+  if (!hits.length) {
+    els.vaultResults.innerHTML = "<p class='muted'>No relevant memory found.</p>";
+    els.vaultResults.classList.remove("hidden");
+    return;
+  }
+  els.vaultResults.innerHTML =
+    `<div class="vault-latest-title">Top ${hits.length} semantic matches</div>` +
+    hits.map((h) =>
+      `<div class="vault-entry"><span class="tag tag-score">${h.score.toFixed(3)}</span>` +
+      `<span class="tag">${escapeHtml(h.source)}</span>` +
+      `<span class="muted mono">${escapeHtml(truncateMiddle(h.text, 140))}</span></div>`
+    ).join("");
+  els.vaultResults.classList.remove("hidden");
+}
+
+async function fetchVault(query) {
+  const qs = query ? `?q=${encodeURIComponent(query)}&top_k=6` : "?top_k=6";
+  const res = await fetch(`/api/vault${qs}`, { cache: "no-store" });
+  const data = await res.json();
+  renderVault(data);
+  if (query) renderVaultSearch(data);
+  return data;
+}
+
 async function runAction(action) {
   const btn = action === "start" ? els.btnStart : els.btnStop;
   const label = btn.textContent;
@@ -280,8 +413,23 @@ els.btnRaw.addEventListener("click", () => {
 els.autoToggle.addEventListener("change", resetAutoTimer);
 els.refreshInterval.addEventListener("change", resetAutoTimer);
 
+els.btnVaultRefresh.addEventListener("click", () => {
+  fetchVault().catch(() => {});
+});
+els.btnVaultSearch.addEventListener("click", () => {
+  const q = els.vaultQuery.value.trim();
+  if (q) fetchVault(q).catch(() => {});
+});
+els.vaultQuery.addEventListener("keydown", (e) => {
+  if (e.key === "Enter") {
+    const q = els.vaultQuery.value.trim();
+    if (q) fetchVault(q).catch(() => {});
+  }
+});
+
 fetchStatus().catch((err) => {
   const msg = err instanceof Error ? err.message : String(err);
   els.rawText.textContent = msg;
 });
+fetchVault().catch(() => {});
 resetAutoTimer();

dashboard/index.html

@@ -87,6 +87,29 @@ <h3>Consensus Head</h3>
       </article>
     </section>
 
+    <section class="panel reveal-5 panel-vault">
+      <div class="panel-head panel-head-wrap">
+        <div>
+          <h3>Memory Vault</h3>
+          <p class="muted" id="vaultSubtitle">Persistent vector memory index</p>
+        </div>
+        <div class="vault-controls">
+          <input id="vaultQuery" class="vault-input" type="text"
+                 placeholder="Semantic search memory… (e.g. pricing, launch)" />
+          <button id="btnVaultSearch" class="btn btn-ghost small">Search</button>
+          <button id="btnVaultRefresh" class="btn btn-ghost small">Refresh</button>
+        </div>
+      </div>
+      <div class="vault-row">
+        <div class="vault-stats" id="vaultStats"></div>
+        <div class="vault-bar-wrap">
+          <canvas id="vaultChart" class="vault-chart" width="520" height="180"></canvas>
+        </div>
+      </div>
+      <div id="vaultLatest" class="vault-latest"></div>
+      <div id="vaultResults" class="vault-results hidden"></div>
+    </section>
+
     <section class="panel reveal-5">
       <div class="panel-head">
         <h3>Recent Log</h3>