ai-helper-agent 2.2.0 → 2.3.1

package/README.md

@@ -2,7 +2,7 @@
 
 # AI Helper Agent
 
-AI Helper Agent — a local proxy service that provides the [AI Helper Chrome Extension](https://github.com/xiweicheng/ai-helper) ([Edge Add-ons](https://microsoftedge.microsoft.com/addons/detail/ai-helper-%E7%BD%91%E9%A1%B5%E6%99%BA%E8%83%BD%E5%8A%A9%E6%89%8B/kabhmgfbkhpbfhhnokaafhkdbckeipcl)) with local file read/write, system command execution, Skill system, MCP protocol extension, and an optional local knowledge base (RAG) capability.
+AI Helper Agent — a local proxy service that provides the [AI Helper Chrome Extension](https://github.com/xiweicheng/ai-helper) ([Chrome Web Store](https://chromewebstore.google.com/detail/ai-helper-%E7%BD%91%E9%A1%B5%E6%99%BA%E8%83%BD%E5%8A%A9%E6%89%8B/fnolhphogjkdcmnecpeckhjoeofldiac) / [Edge Add-ons](https://microsoftedge.microsoft.com/addons/detail/ai-helper-%E7%BD%91%E9%A1%B5%E6%99%BA%E8%83%BD%E5%8A%A9%E6%89%8B/kabhmgfbkhpbfhhnokaafhkdbckeipcl)) with local file read/write, system command execution, Skill system, MCP protocol extension, and an optional local knowledge base (RAG) capability.
 
 ## Installation
 
@@ -156,6 +156,7 @@ Configuration file path: `~/.ai-helper-agent/config.json`
 | `pairCodeTTL` | Pairing code validity (seconds) | 30 |
 | `commandTimeout` | Command execution timeout (ms) | 300000 (5 min) |
 | `fileMaxSize` | Max file read/write size (bytes) | 52428800 (50 MB) |
+| `memoryEmbedding` | Semantic memory recall channel config — see [Semantic recall channel](#semantic-recall-channel-tiers). Managed through `/api/memory/embedding-config`; `enabled: false` keeps recall keyword-only | disabled |
 
 ## File Search
 
@@ -243,6 +244,15 @@ Results are returned in reverse chronological order (newest first).
 | POST | `/api/fs/delete` | Delete file/directory |
 | POST | `/api/fs/search_files` | Search files by name pattern (glob) |
 | POST | `/api/fs/search_content` | Search file content (rg preferred, Node.js fallback) |
+| POST | `/api/memory/store` | Write long-term memory (action: add / update / delete) |
+| POST | `/api/memory/recall` | Recall long-term memory by keyword scoring (with per-session dedup) |
+| POST | `/api/memory/manage` | Memory review (`action: review`) / compaction (`action: compact`) |
+| POST | `/api/memory/inject` | Three-tier injection selection for the system prompt |
+| POST | `/api/memory/reindex` | Rebuild the memory vector cache for every entry using the current embedding config |
+| GET | `/api/memory/embedding-config` | Read the semantic recall channel config (plain key never echoed; `hasApiKey` included) |
+| POST | `/api/memory/embedding-config` | Save the semantic recall channel config |
+| POST | `/api/memory/test-embedding` | Probe the embedding endpoint (returns vector dimensions on success) |
+| GET | `/api/memory/vector-status` | Read-only vector index status: `indexed` / `total`, stored vs current fingerprint (`storedModelKey` / `currentModelKey`), `stale`, `tier`. Never builds a provider, so querying it cannot trigger a model download |
 | POST | `/api/exec` | Execute system command |
 | POST | `/api/exec/stop` | Stop command execution |
 | GET | `/api/exec/running` | List running processes |
@@ -381,6 +391,89 @@ Configuration file path: `~/.ai-helper-agent/config.json`
 4. Tool call requests are forwarded to MCP servers via the agent
 5. Tool results are returned to the extension
 
+## Long-term Memory
+
+Cross-session long-term memory is owned by the agent: the storage file, the read-modify-write lock, recall scoring, review/compaction and injection selection all run in this process. The extension only forwards tool calls to the endpoints below — no generic file-API round-trips, no dependency on the extension Service Worker staying alive.
+
+- **Data location** — `~/.ai-helper-agent/memory/global-memory.json` (facts + summaries + stats)
+- **Concurrency** — an in-process promise-chain lock serializes every read-modify-write; the per-session recall cache (which memories were already surfaced in this conversation) lives in agent memory
+- **i18n** — response text is rendered from the request's `Accept-Language`
+- **Where summaries come from** — besides explicit tool writes, `type: "summary"` entries are sedimented automatically by the extension's session context compaction (tags `compaction` + `session:<id>`, `importance: 6`, `category: "decision"`); re-compacting the same session updates the same `memoryId` instead of piling up
+
+### Record schema
+
+```json
+{
+  "id": "fact_a1b2c3d4_xxxx",
+  "type": "fact | summary",
+  "category": "custom",
+  "content": "...",
+  "tags": ["..."],
+  "importance": 5,
+  "accessCount": 0,
+  "lastAccessAt": null,
+  "createdAt": "2026-01-15T10:30:00.000Z",
+  "updatedAt": "2026-01-15T10:30:00.000Z",
+  "sourceSessionId": null
+}
+```
+
+### Recall scoring
+
+A memory's keyword score is `Σ (matched units in a field × that field's weight)`, plus whole-query bonuses:
+
+| Signal | Score |
+|--------|-------|
+| matched unit in `content` | × 3 per unit |
+| matched unit in `title` | × 2 per unit |
+| matched unit in `tags` | × 2 per unit |
+| full query (≥ 2 chars) found in `content` | +5 |
+| full query (≥ 2 chars) found in `title` | +3 |
+
+Units come from the **longest matching fragment**, not from chopping the query into pieces: a CJK fragment of length 2 counts as 1 unit, 3 counts as 2, ≥ 4 counts as 3 (top of the scale), and each whole latin/digit word counts as 1 — capped at 3 units per field. Extraction and matching are zero-dependency and **shared with knowledge-base retrieval** (`rag/keyword-match.js`), so a query means the same thing in both subsystems.
+
+Two gates keep irrelevant memories out:
+
+- **Document frequency** — a unit found in over half of the candidates (and in more than 2 of them) carries no discriminating power and scores nothing.
+- **Hit threshold** — a memory counts as a hit only when it matched a *whole* unit (an entire latin word or an entire CJK run) or at least 2 units. Sharing only a random 2-character fragment is no longer a hit; the exact-match bonuses alone cannot rescue such an entry.
+
+### Semantic recall channel (tiers)
+
+A vector channel can be layered on top of keyword scoring. It runs at one of three tiers and **the default is always Tier A**:
+
+| Tier | Condition | Behaviour |
+|------|-----------|-----------|
+| A `none` | Unconfigured or `enabled: false` (default) | Keyword scoring only, exactly as above; no new network calls |
+| B `remote` | `mode: "openai-compat"` with both apiKey and modelName set | Remote embeddings + in-process cosine comparison |
+| C `local` | `mode: "local"` and RAG dependencies present locally | Vectors from a local transformers model |
+
+- **Vectors are a derived cache** — `~/.ai-helper-agent/memory/memory-vectors.json` can be deleted at any time; `global-memory.json` is the single source of truth. A missing or corrupt cache is treated as empty and refilled on demand during recall — it never throws.
+- **Changing the model invalidates everything** — the cache key is `mode|modelName|dimensions`; if any part changes the old vectors are dropped rather than compared (mismatched dimensions cannot be compared). No manual cleanup needed.
+- **Failure degrades, it does not break memory** — when the embedding endpoint errors or exceeds its 8s timeout, that recall falls back to Tier A and still answers 200. `tier` and `vectorMatched` in the recall response say which tier actually ran and how many hits were accepted by the semantic route (capped at 2).
+- **Both channels fuse ranks, not scores** — keywords and vectors each produce their own ranking, which is merged as `1/(60 + rank)` (RRF). The two scales are not comparable, so only **rank** is consumed, never magnitude: an item hit by both routes naturally floats to the top, one hit by a single route lands behind it.
+- **The semantic gate has two stages, not one absolute cosine threshold** — (1) **query-level floor**: if the best cosine of this query is below `0.48`, the library is judged to hold no answer for it and the whole semantic route contributes nothing (recall output goes straight into the prompt — "everything is marginally relevant" equals "nothing is relevant"); (2) **item-level relative drop-off**: only items within `0.05` of that best score are accepted, at most 2.
+- **Why there is no absolute similarity threshold** — measured on a real 41-item library: absolute cosine fails at item level (irrelevant queries reach top1 scores of 0.3340~0.4506, while a ≥0.3 threshold admits 40/41 items and ≥0.45 still admits 20/41). It separates only at query-level top1 (irrelevant 0.3340~0.4506 vs relevant 0.5467~0.6353, gap 0.096), which is exactly where the floor above is applied.
+- **Known limits** — the `0.48` floor was calibrated on `BAAI/bge-m3` with a 41-item library and 9 queries; its usable window is `(0.4506, 0.5163]` (roughly 0.03 of headroom on each side). **Recalibration is mandatory when the embedding model or the memory-text concatenation changes**, and the rule needs re-measurement once the library grows to a few hundred items. Separation (`top1 − top2`) was the original main gate of this channel and was falsified by measurement: the larger the library, the more an irrelevant query's top1−top2 can grow by chance, and the real library's target query had a gap of 0.0214 — below two irrelevant queries at 0.0361 / 0.0397, so the two are inseparable on that axis. **Do not add it back.** Practical consequence: when a target item is not the library's best cosine and falls outside the `0.05` drop-off, the semantic route will not return it (on the real library such items still enter via the keyword route, at a lower rank).
+- **The local backend only indexes the first 512 tokens (the tail is dropped silently)** — the tokenizer window of `Xenova/bge-small-zh-v1.5` is 512 tokens and inference cuts the tail **without any error** (measured: 13 of 41 items in a real library exceed it; the largest one gets only 46% of its content in). Blast radius is limited: the semantic channel cannot see the tail, while **the keyword channel still reads the full text**. The knowledge base is unaffected (its default `chunkSize: 400` already fits). Remote models have far larger windows, and a vendor's truncation behaviour is invisible in the vectors it returns — so the remote tier reports nothing. This round changes no vector content; it only makes the overflow visible: `reindex` now returns `overWindow: { count, window, ids }` and the process log names the items.
+- **Vectors drift slightly with batch composition (known noise, not a bug)** — embedding one memory alone vs inside a batch gives cos 0.9940–0.9974 (measured), caused by the `q8` quantized kernels and unrelated to how long the batch partners are; no observable effect on rank fusion and no effect on recall results.
+- **Operations entry point** — after first enabling the channel or switching models, call `POST /api/memory/reindex` once to backfill existing memories (it returns `skipped` when the channel is off, and a clear 400 — not 500 — when the endpoint is unreachable). When only some batches fail the response still reports success but carries `failedBatches` plus `lastError` (the first reason); the full reasons go to the process log. On the local tier it also carries `overWindow` (how many items exceed the window, and which).
+- **Configuration UI** — the extension's options page has a **Memory** tab that reads and writes this channel (`/api/memory/embedding-config`), probes the endpoint (`/api/memory/test-embedding`, remote tier only — the local tier's first call has to load the model and would always blow the 8s budget, so the button is hidden there instead of lying), rebuilds the cache (`/api/memory/reindex` behind an explicit confirmation that quotes the measured cost) and shows `GET /api/memory/vector-status`, whose `stale` flag is what makes "the model changed but nobody rebuilt" visible instead of silently degrading recall. Note the tab is a convenience, not a restart workaround: `loadConfig()` caches by file mtime and the provider cache is keyed by the fingerprint, so editing `config.json` by hand takes effect just as immediately. Its real value is discoverability, verification and mis-fill prevention. One rule the form follows strictly — a field the user cannot see is **absent from the request body**, never empty or `0`: sending `dimensions: 0` would overwrite a stored `512`, change the fingerprint, raise a false `stale` and force a pointless rebuild.
+- **Fingerprint staleness is report-only** — `vector-status` tells you the stored vectors were built by a different `mode|modelName|dimensions`; it does not discard or rewrite them, because `stale` with `indexed > 0` is a state worth looking at (and an empty library is never called stale — first-time enablement is normal).
+
+### Injection tiers (system prompt)
+
+| Tier | Rule |
+|------|------|
+| 1 — resident | `importance >= 8` is always injected |
+| 2 — budgeted | ranked by value, capped at 1500 characters and 12 items |
+| 3 — index only | remaining memories surface as tag index lines; the model pulls them with a recall call |
+
+> Entries with `type: "summary"` are **never** part of Tier 1/2 injection — they count only toward Tier 3 (tag index + on-demand recall), so compaction summaries never eat system-prompt budget; the model pulls them when needed.
+
+Memory value = `importance x (1 + ln(accessCount + 1)) x timeDecay`; entries whose value falls below 10 are evicted by compaction (`/api/memory/manage` with `action: "compact"`).
+
+> Long-term memory needs an agent build that ships the `/api/memory/*` endpoints. The extension deliberately does **not** fall back to the generic file API — a missing endpoint is reported as a version-mismatch hint instead.
+
 ## Knowledge Base (RAG)
 
 Optional local retrieval-augmented generation. Turn documents into searchable knowledge bases so the extension can answer with cited sources. The RAG stack ships as optional dependencies — the agent detects availability at startup and reports it via `/api/rag/status`; when dependencies are missing, RAG requests return `503` and the extension offers one-click installation (whitelisted packages, progress polled from `/api/rag/install/status`). Installation is also available from the CLI: `ai-helper-agent rag status` / `ai-helper-agent rag install`. Users who have installed the dependencies successfully before get **automatic recovery on startup**: `npm -g` upgrades prune manually-installed packages inside the package directory, so the agent re-installs them automatically (marked `trigger=auto`) — no manual action needed; users who never installed successfully are not prompted.

package/README.zh-CN.md

@@ -2,7 +2,7 @@
 
 # AI Helper Agent
 
-AI Helper 代理服务，为 [AI Helper Chrome 扩展](https://github.com/xiweicheng/ai-helper)（[Edge 扩展商店](https://microsoftedge.microsoft.com/addons/detail/ai-helper-%E7%BD%91%E9%A1%B5%E6%99%BA%E8%83%BD%E5%8A%A9%E6%89%8B/kabhmgfbkhpbfhhnokaafhkdbckeipcl)）提供本地文件读写、系统命令执行、Skill 技能系统、MCP 协议扩展与可选本地知识库（RAG）能力。
+AI Helper 代理服务，为 [AI Helper Chrome 扩展](https://github.com/xiweicheng/ai-helper)（[Chrome 应用商店](https://chromewebstore.google.com/detail/ai-helper-%E7%BD%91%E9%A1%B5%E6%99%BA%E8%83%BD%E5%8A%A9%E6%89%8B/fnolhphogjkdcmnecpeckhjoeofldiac) / [Edge 扩展商店](https://microsoftedge.microsoft.com/addons/detail/ai-helper-%E7%BD%91%E9%A1%B5%E6%99%BA%E8%83%BD%E5%8A%A9%E6%89%8B/kabhmgfbkhpbfhhnokaafhkdbckeipcl)）提供本地文件读写、系统命令执行、Skill 技能系统、MCP 协议扩展与可选本地知识库（RAG）能力。
 
 ## 安装
 
@@ -156,6 +156,7 @@ Agent 通过 PID 文件（`~/.ai-helper-agent/agent.pid`）管理进程生命周
 | `pairCodeTTL` | 配对码有效期（秒） | 30 |
 | `commandTimeout` | 命令执行超时（毫秒） | 300000（5分钟） |
 | `fileMaxSize` | 文件读写最大字节数 | 52428800（50MB） |
+| `memoryEmbedding` | 记忆语义召回通道配置，详见[语义召回通道](#语义召回通道tier)；推荐通过 `/api/memory/embedding-config` 读写，`enabled: false` 时 recall 只走关键词 | 关闭 |
 
 ## 文件搜索
 
@@ -243,6 +244,15 @@ GET /api/logs/dates
 | POST | `/api/fs/delete` | 删除文件/目录 |
 | POST | `/api/fs/search_files` | 按文件名模式搜索（glob） |
 | POST | `/api/fs/search_content` | 搜索文件内容（rg 优先，Node.js 回退） |
+| POST | `/api/memory/store` | 写入长期记忆（action：add / update / delete） |
+| POST | `/api/memory/recall` | 关键词计分召回长期记忆（含会话级去重） |
+| POST | `/api/memory/manage` | 记忆审查（`action: review`）/ 压缩淘汰（`action: compact`） |
+| POST | `/api/memory/inject` | 为系统提示词执行三级注入筛选 |
+| POST | `/api/memory/reindex` | 按当前 embedding 配置重建全部记忆向量缓存 |
+| GET | `/api/memory/embedding-config` | 读取语义召回通道配置（明文 key 永不回显，附 `hasApiKey`） |
+| POST | `/api/memory/embedding-config` | 保存语义召回通道配置 |
+| POST | `/api/memory/test-embedding` | 探活 embedding 端点（成功回报向量维度） |
+| GET | `/api/memory/vector-status` | 只读向量索引状态：`indexed` / `total`、已存指纹与当前指纹（`storedModelKey` / `currentModelKey`）、`stale`、`tier`。不构造 provider，因此查状态不会触发模型下载 |
 | POST | `/api/exec` | 执行系统命令 |
 | POST | `/api/exec/stop` | 停止命令执行 |
 | GET | `/api/exec/running` | 运行中的进程列表 |
@@ -381,6 +391,89 @@ enabled: true
 4. 工具调用请求通过 Agent 转发到 MCP Server
 5. 工具结果返回给扩展
 
+## 长期记忆
+
+跨会话长期记忆由 Agent 全权托管：存储文件、读写互斥锁、召回计分、审查/淘汰、注入筛选全部在本进程内完成。插件只向下表端点转发工具调用——不再走通用文件 API 往返，也不依赖插件 Service Worker 是否存活。
+
+- **数据位置** — `~/.ai-helper-agent/memory/global-memory.json`（facts + summaries + stats）
+- **并发** — 进程内 Promise 链互斥锁串行化每一次读-改-写；会话级召回缓存（本对话已呈现过哪些记忆）常驻 Agent 内存
+- **国际化** — 响应文案根据请求 `Accept-Language` 渲染
+- **摘要来源** — 除工具主动写入外，`type: "summary"` 条目可由插件的会话上下文压缩自动沉淀（标签 `compaction` + `session:<会话 ID>`，`importance: 6`，`category: "decision"`）；同一会话重复压缩按 `memoryId` 更新同一条，不堆积
+
+### 记录结构
+
+```json
+{
+  "id": "fact_a1b2c3d4_xxxx",
+  "type": "fact | summary",
+  "category": "custom",
+  "content": "...",
+  "tags": ["..."],
+  "importance": 5,
+  "accessCount": 0,
+  "lastAccessAt": null,
+  "createdAt": "2026-01-15T10:30:00.000Z",
+  "updatedAt": "2026-01-15T10:30:00.000Z",
+  "sourceSessionId": null
+}
+```
+
+### 召回计分
+
+记忆关键词分 = `Σ（某字段上的命中单元数 × 该字段权重）`，另加整句原样命中分：
+
+| 命中信号 | 分值 |
+|----------|------|
+| `content` 的命中单元 | × 3 / 单元 |
+| `title` 的命中单元 | × 2 / 单元 |
+| `tags` 的命中单元 | × 2 / 单元 |
+| 查询原文（≥ 2 字）出现在 `content` 中 | +5 |
+| 查询原文（≥ 2 字）出现在 `title` 中 | +3 |
+
+单元数来自**最长命中片段**，而不是把查询拆碎：中文片段长 2 计 1 个单元、长 3 计 2 个、≥ 4 计 3 个（顶格），英文/数字整词每命中一个计 1 个；单字段封顶 3 个单元。切段与匹配零依赖，且**与知识库检索共用同一份实现**（`rag/keyword-match.js`），同一条查询在两个子系统里含义一致。
+
+两道闸门挡住无关条目：
+
+- **文档频率剔除** —— 在过半候选（且超过 2 条）里出现的单元无区分度，不参与计分。
+- **命中门槛** —— 只有命中了**完整单元**（整个英文词或整个中文连续段）或累计 ≥ 2 个单元才算命中。只共享一个随机 2 字碎片不再算命中，整句原样加分也不能单独救回这样的条目。
+
+### 语义召回通道（Tier）
+
+关键词之外可以叠加一路向量召回，分三档，**缺省永远是 A 档**：
+
+| 档位 | 条件 | 行为 |
+|------|------|------|
+| A `none` | 未配置或 `enabled: false`（缺省） | 只走上表的关键词计分，不产生任何新的网络调用 |
+| B `remote` | `mode: "openai-compat"` 且 apiKey 与 modelName 都已填 | 远端 embedding + 进程内余弦比对 |
+| C `local` | `mode: "local"` 且本机已具备 RAG 依赖 | 本地 transformers 模型出向量 |
+
+- **向量是派生缓存** — `~/.ai-helper-agent/memory/memory-vectors.json` 可随时整删重建；`global-memory.json` 是唯一真源。缓存缺失或损坏一律当空处理，并在召回命中时按需补齐，绝不抛错。
+- **模型一变即整体作废** — 缓存键是 `mode|modelName|dimensions`，任一变化都不再复用旧向量（维度不同无法比对），无需手工清理。
+- **失败即降级，不拖垮记忆** — embedding 端点报错或超时（上限 8s）时，该次召回自动退回 A 档且响应仍是 200。`recall` 响应里的 `tier` 与 `vectorMatched` 说明这次实际走了哪一档、有几条被语义路采信（后者上限 2）。
+- **双通道只融合名次，不融合分数** — 关键词与向量各排各的名次，按 `1/(60 + 名次)` 相加后重排（RRF）。两路分值刻度不可比，因此**只比名次、不比大小**：同时被两路命中的条目自然排前，只被一路命中的靠后。
+- **语义采信门是两道，而不是「cos 超过某个绝对阈值」** — ① **查询级落点地板**：本次查询的最高 cos 低于 `0.48` 时判定「本库对这个查询没有答案」，整个语义通道一条都不采信（记忆会直接注入提示词，「全都勉强相关」等于「没有相关」）；② **条目级相对落差**：只采信距该最高分 `0.05` 以内的头部，最多 2 条。
+- **为什么没有绝对相似度阈值** — 在真实 41 条库上：条目级绝对 cos 不可用（无关查询的 top1 落在 0.3340~0.4506，而阈值 ≥0.3 会放过 40/41 条、≥0.45 仍放过 20/41）；它只在查询级 top1 上可分（无关 0.3340~0.4506 vs 相关 0.5467~0.6353，间隙 0.096），所以上述地板只作用在查询级。
+- **已知限制** — 地板 `0.48` 在 `BAAI/bge-m3` + 41 条库 + 9 条查询上标定，可用窗口为 `(0.4506, 0.5163]`（上下余量各约 0.03），**换 embedding 模型或改变记忆文本拼接方式后必须重新标定**；记忆条目数增长到几百条后该规则也需重测。分离度（`top1 − top2`）曾是本通道的主门，已被实测否定：库越大、无关查询的 top1−top2 越可能偶然变大，而真实库目标查询的 gap（0.0214）低于两条无关查询的（0.0361 / 0.0397），在 gap 轴上根本分不开——**不要把它加回来**。具体后果：目标条目的 cos 不是全场第 1、且落在头部 `0.05` 落差之外时，语义路不会返回它（真实库上这种条目靠词面路进窗口，名次会被压低）。
+- **本地档只索引前 512 token（超出部分静默不进向量）** — 本地 `Xenova/bge-small-zh-v1.5` 的 tokenizer 窗口为 512 token，推理时**不报错地**砍掉尾部（实测真实库 41 条中 13 条超窗，最大一条只吃进 46% 的内容）。影响面有限：语义路看不到尾部，而**词面路拿的是全文**。知识库侧不受影响（默认 `chunkSize: 400` 已在窗口内）。远端模型窗口普遍大得多，且厂商截断行为从返回向量里看不出来，所以远端档不报该字段。本轮不改向量内容，只把超窗变成可见：`reindex` 响应带 `overWindow: { count, window, ids }`，进程日志点名是哪几条。
+- **向量会随分批组合轻微漂移（已知噪声，非缺陷）** — 同一条记忆逐条嵌 vs 同批嵌的 cos 在 0.9940~0.9974（实测），源于 `q8` 量化核的数值差异，与批内文本长短无关；对名次融合无可观测影响，不影响召回结果。
+- **运维入口** — 首次启用或更换模型后调一次 `POST /api/memory/reindex` 把存量记忆补进缓存（未启用时返回 `skipped`，端点不可达则明确 400 而非 500）。部分批失败时响应仍为成功但带 `failedBatches` 与 `lastError`（首条原因），全量原因进进程日志；本地档还会带 `overWindow`（超窗条数与点名）。
+- **配置界面** — 扩展 options 页新增「记忆」Tab：读写本通道（`/api/memory/embedding-config`）、探活（`/api/memory/test-embedding`，**仅远端档**——本地档首次调用要先加载模型，必定击穿 8s 超时，所以那里直接隐藏按钮而不是给个骗人的失败）、重建缓存（`/api/memory/reindex`，执行前弹确认并写明实测耗时）、展示 `GET /api/memory/vector-status`——其中 `stale` 才是把「改了模型但没人重建」从静默退化变成可见的那个信号。注意这个界面不是为「免重启」而做的：`loadConfig()` 按文件 mtime 缓存、provider 缓存又按指纹取键，手改 `config.json` 同样即时生效。它的真实价值是可发现、可验证、防填错。表单硬守一条规则——用户看不见的字段从请求体里**缺席**，绝不交空值或 `0`：交 `dimensions: 0` 会把已存的 512 覆盖掉、改变指纹、冒报 `stale`，逼用户做一次无谓重建。
+- **指纹陈旧只报告不改数据** — `vector-status` 只告诉你已存向量是由另一个 `mode|modelName|dimensions` 生成的，不删除也不重写：`stale` 且 `indexed > 0` 本身就是要看到的状态（而空库永不判为陈旧——首次启用是正常态）。
+
+### 三级注入策略（系统提示词）
+
+| 层级 | 规则 |
+|------|------|
+| 1 — 常驻 | `importance >= 8` 总是注入 |
+| 2 — 预算 | 按价值排序，上限 1500 字符 + 12 条 |
+| 3 — 仅索引 | 其余记忆只以标签索引行呈现，由模型主动 recall 拉取 |
+
+> `type: "summary"` 的条目**不参与** Tier 1/2 注入，只计入 Tier 3（标签索引 + 按需召回）——会话压缩摘要因此不会挤占系统提示词预算，需要时由模型自己拉取。
+
+记忆价值 = `importance x (1 + ln(accessCount + 1)) x 时间衰减`；价值降到 10 以下的条目由压缩淘汰（`/api/memory/manage` 的 `action: "compact"`）。
+
+> 长期记忆需要带 `/api/memory/*` 端点的 Agent 版本。插件**不会**回退到通用文件 API——端点缺失时直接提示版本不匹配。
+
 ## 知识库（RAG）
 
 可选的本地检索增强生成（RAG）能力：将文档索引为可检索的知识库，让插件回答时附带引用来源。RAG 依赖为可选安装 —— Agent 启动时自动检测可用性，并通过 `/api/rag/status` 对外暴露；依赖缺失时 RAG 请求返回 `503`，插件端提供一键安装（白名单依赖，进度可经 `/api/rag/install/status` 轮询）。也可通过 CLI 安装：`ai-helper-agent rag status` / `ai-helper-agent rag install`。曾成功安装过依赖的用户享受**启动自动恢复**：`npm -g` 更新代理会清理包目录内手动安装的依赖，Agent 启动检测到缺失时自动重新安装（标记 `trigger=auto`），用户零操作；从未安装成功过的用户不触发。

package/dist/assets/{agent-store-DOyMdB_7.js → agent-store-DcAs6YyA.js}

@@ -1,2 +1,2 @@
-import{r as e}from"./rolldown-runtime-C0FnF6B9.js";import{a as t,c as n}from"./i18n-MmYITlVj.js";import{_ as r,g as i}from"./session-store-B0ZPGm7d.js";var a=e({createAgent:()=>m,deleteAgent:()=>g,getActiveAgentId:()=>_,getAgent:()=>p,getAllAgents:()=>d,getCustomAgents:()=>f,setActiveAgentId:()=>v,updateAgent:()=>h});t(`zh`,{agentStore:{unnamedAgent:`未命名Agent`}}),t(`en`,{agentStore:{unnamedAgent:`Unnamed Agent`}});var o=`customAgents`,s=`activeAssistantId`,c=`activeAgentId`,l=!1;async function u(){if(l)return;l=!0;let e=await chrome.storage.local.get([c,s]),t=e[c],n=e[s];t&&!n&&!t.startsWith(`pa_`)&&(await chrome.storage.local.set({[s]:t}),await chrome.storage.local.remove(c))}async function d(){let e=(await chrome.storage.local.get([o]))[o]||[];return[...i,...e]}async function f(){return(await chrome.storage.local.get([o]))[o]||[]}async function p(e){return e&&(await d()).find(t=>t.id===e)||null}async function m(e){let t=await f(),i={id:r(),name:e.name||n(`agentStore.unnamedAgent`),description:e.description||``,icon:e.icon||`🤖`,systemPrompt:e.systemPrompt||``,toolIds:e.toolIds||null,skillIds:e.skillIds||null,mcpExcludedServerIds:e.mcpExcludedServerIds??null,isBuiltin:!1,allowSubDispatch:e.allowSubDispatch!==void 0&&e.allowSubDispatch,model:e.model||null,temperature:e.temperature===void 0?null:e.temperature,topP:e.topP===void 0?null:e.topP,createdAt:new Date().toISOString(),updatedAt:new Date().toISOString()};return t.push(i),await chrome.storage.local.set({[o]:t}),i}async function h(e,t){let n=await f(),r=n.findIndex(t=>t.id===e);return r===-1?null:(n[r]={...n[r],...t,id:e,isBuiltin:!1,updatedAt:new Date().toISOString()},await chrome.storage.local.set({[o]:n}),n[r])}async function g(e){let t=await f(),n=t.findIndex(t=>t.id===e);return n!==-1&&(t.splice(n,1),await chrome.storage.local.set({[o]:t}),(await chrome.storage.local.get([s]))[s]===e&&await chrome.storage.local.remove(s),!0)}async function _(){return await u(),(await chrome.storage.local.get([s]))[s]||null}async function v(e){e?await chrome.storage.local.set({[s]:e}):await chrome.storage.local.remove(s)}export{p as a,h as c,_ as i,m as n,d as o,g as r,v as s,a as t};
-//# sourceMappingURL=agent-store-DOyMdB_7.js.map
+import{r as e}from"./rolldown-runtime-C0FnF6B9.js";import{a as t,c as n}from"./i18n-C4jSRoCH.js";import{_ as r,v as i}from"./session-store-CzVQCgQ_.js";var a=e({createAgent:()=>m,deleteAgent:()=>g,getActiveAgentId:()=>_,getAgent:()=>p,getAllAgents:()=>d,getCustomAgents:()=>f,setActiveAgentId:()=>v,updateAgent:()=>h});t(`zh`,{agentStore:{unnamedAgent:`未命名Agent`}}),t(`en`,{agentStore:{unnamedAgent:`Unnamed Agent`}});var o=`customAgents`,s=`activeAssistantId`,c=`activeAgentId`,l=!1;async function u(){if(l)return;l=!0;let e=await chrome.storage.local.get([c,s]),t=e[c],n=e[s];t&&!n&&!t.startsWith(`pa_`)&&(await chrome.storage.local.set({[s]:t}),await chrome.storage.local.remove(c))}async function d(){let e=(await chrome.storage.local.get([o]))[o]||[];return[...r,...e]}async function f(){return(await chrome.storage.local.get([o]))[o]||[]}async function p(e){return e&&(await d()).find(t=>t.id===e)||null}async function m(e){let t=await f(),r={id:i(),name:e.name||n(`agentStore.unnamedAgent`),description:e.description||``,icon:e.icon||`🤖`,systemPrompt:e.systemPrompt||``,toolIds:e.toolIds||null,skillIds:e.skillIds||null,mcpExcludedServerIds:e.mcpExcludedServerIds??null,isBuiltin:!1,allowSubDispatch:e.allowSubDispatch!==void 0&&e.allowSubDispatch,model:e.model||null,temperature:e.temperature===void 0?null:e.temperature,topP:e.topP===void 0?null:e.topP,createdAt:new Date().toISOString(),updatedAt:new Date().toISOString()};return t.push(r),await chrome.storage.local.set({[o]:t}),r}async function h(e,t){let n=await f(),r=n.findIndex(t=>t.id===e);return r===-1?null:(n[r]={...n[r],...t,id:e,isBuiltin:!1,updatedAt:new Date().toISOString()},await chrome.storage.local.set({[o]:n}),n[r])}async function g(e){let t=await f(),n=t.findIndex(t=>t.id===e);return n!==-1&&(t.splice(n,1),await chrome.storage.local.set({[o]:t}),(await chrome.storage.local.get([s]))[s]===e&&await chrome.storage.local.remove(s),!0)}async function _(){return await u(),(await chrome.storage.local.get([s]))[s]||null}async function v(e){e?await chrome.storage.local.set({[s]:e}):await chrome.storage.local.remove(s)}export{p as a,h as c,_ as i,m as n,d as o,g as r,v as s,a as t};
+//# sourceMappingURL=agent-store-DcAs6YyA.js.map

package/dist/assets/{agent-store-DOyMdB_7.js.map → agent-store-DcAs6YyA.js.map}

@@ -1 +1 @@
-{"version":3,"file":"agent-store-DOyMdB_7.js","names":[],"sources":["../../src/side_panel/agent-store.js"],"sourcesContent":["// side_panel/agent-store.js - Agent 数据持久化\nimport { BUILTIN_AGENTS, generateAgentId } from '../shared/agent-defaults.js';\nimport { t, registerTranslations } from '../shared/i18n.js';\n\nregisterTranslations('zh', {\n  agentStore: {\n    unnamedAgent: '未命名Agent',\n  },\n});\nregisterTranslations('en', {\n  agentStore: {\n    unnamedAgent: 'Unnamed Agent',\n  },\n});\n\nconst STORAGE_KEY = 'customAgents';\nconst ACTIVE_KEY = 'activeAssistantId';\nconst LEGACY_ACTIVE_KEY = 'activeAgentId';\n\nlet migratedActiveKey = false;\n\nasync function migrateLegacyActiveKey() {\n  if (migratedActiveKey) return;\n  migratedActiveKey = true;\n\n  const result = await chrome.storage.local.get([LEGACY_ACTIVE_KEY, ACTIVE_KEY]);\n  const legacyValue = result[LEGACY_ACTIVE_KEY];\n  const newValue = result[ACTIVE_KEY];\n\n  if (legacyValue && !newValue && !legacyValue.startsWith('pa_')) {\n    await chrome.storage.local.set({ [ACTIVE_KEY]: legacyValue });\n    await chrome.storage.local.remove(LEGACY_ACTIVE_KEY);\n  }\n}\n\n/**\n * 获取所有 Agent（内置 + 用户自定义）\n * @returns {Promise<Array>}\n */\nexport async function getAllAgents() {\n  const result = await chrome.storage.local.get([STORAGE_KEY]);\n  const customAgents = result[STORAGE_KEY] || [];\n  return [...BUILTIN_AGENTS, ...customAgents];\n}\n\n/**\n * 仅获取用户自定义的 Agent\n * @returns {Promise<Array>}\n */\nexport async function getCustomAgents() {\n  const result = await chrome.storage.local.get([STORAGE_KEY]);\n  return result[STORAGE_KEY] || [];\n}\n\n/**\n * 根据 ID 获取单个 Agent\n * @param {string} agentId\n * @returns {Promise<Object|null>}\n */\nexport async function getAgent(agentId) {\n  if (!agentId) return null;\n  const all = await getAllAgents();\n  return all.find(a => a.id === agentId) || null;\n}\n\n/**\n * 创建新的自定义 Agent\n * @param {Object} agentData - 不含 id 的 Agent 数据\n * @returns {Promise<Object>}\n */\nexport async function createAgent(agentData) {\n  const agents = await getCustomAgents();\n  const newAgent = {\n    id: generateAgentId(),\n    name: agentData.name || t('agentStore.unnamedAgent'),\n    description: agentData.description || '',\n    icon: agentData.icon || '🤖',\n    systemPrompt: agentData.systemPrompt || '',\n    toolIds: agentData.toolIds || null,  // null = 继承全局\n    skillIds: agentData.skillIds || null, // null = 继承全部启用技能\n    mcpExcludedServerIds: agentData.mcpExcludedServerIds ?? null, // MCP 服务排除列表（服务级 deny-list）\n    isBuiltin: false,\n    allowSubDispatch: agentData.allowSubDispatch !== undefined ? agentData.allowSubDispatch : false,\n    model: agentData.model || null,\n    temperature: agentData.temperature !== undefined ? agentData.temperature : null,\n    topP: agentData.topP !== undefined ? agentData.topP : null,\n    createdAt: new Date().toISOString(),\n    updatedAt: new Date().toISOString(),\n  };\n  agents.push(newAgent);\n  await chrome.storage.local.set({ [STORAGE_KEY]: agents });\n  return newAgent;\n}\n\n/**\n * 更新已有的自定义 Agent\n * @param {string} agentId\n * @param {Object} updates\n * @returns {Promise<Object|null>}\n */\nexport async function updateAgent(agentId, updates) {\n  const agents = await getCustomAgents();\n  const idx = agents.findIndex(a => a.id === agentId);\n  if (idx === -1) return null;\n  \n  agents[idx] = {\n    ...agents[idx],\n    ...updates,\n    id: agentId,  // id 不可变\n    isBuiltin: false,  // 不可变\n    updatedAt: new Date().toISOString(),\n  };\n  await chrome.storage.local.set({ [STORAGE_KEY]: agents });\n  return agents[idx];\n}\n\n/**\n * 删除自定义 Agent\n * @param {string} agentId\n * @returns {Promise<boolean>}\n */\nexport async function deleteAgent(agentId) {\n  const agents = await getCustomAgents();\n  const idx = agents.findIndex(a => a.id === agentId);\n  if (idx === -1) return false;\n  \n  agents.splice(idx, 1);\n  await chrome.storage.local.set({ [STORAGE_KEY]: agents });\n\n  // 如果删除的 Agent 是当前选中的，重置为默认\n  const result = await chrome.storage.local.get([ACTIVE_KEY]);\n  if (result[ACTIVE_KEY] === agentId) {\n    await chrome.storage.local.remove(ACTIVE_KEY);\n  }\n  \n  return true;\n}\n\n/**\n * 获取当前选中的 Agent ID\n * @returns {Promise<string|null>}\n */\nexport async function getActiveAgentId() {\n  await migrateLegacyActiveKey();\n  const result = await chrome.storage.local.get([ACTIVE_KEY]);\n  return result[ACTIVE_KEY] || null;\n}\n\n/**\n * 设置当前选中的 Agent ID\n * @param {string|null} agentId\n */\nexport async function setActiveAgentId(agentId) {\n  if (agentId) {\n    await chrome.storage.local.set({ [ACTIVE_KEY]: agentId });\n  } else {\n    await chrome.storage.local.remove(ACTIVE_KEY);\n  }\n}\n\n/**\n * 从模板创建 Agent\n * @param {Object} template\n * @returns {Promise<Object>}\n */\nexport async function createAgentFromTemplate(template) {\n  return createAgent({\n    name: template.name,\n    icon: template.icon,\n    description: template.description,\n    systemPrompt: template.systemPrompt,\n    toolIds: template.toolIds ? [...template.toolIds] : null,\n    allowSubDispatch: template.allowSubDispatch || false,\n  });\n}\n"],"mappings":"+TAIA,EAAqB,KAAM,CACzB,WAAY,CACV,aAAc,UAChB,CACF,CAAC,EACD,EAAqB,KAAM,CACzB,WAAY,CACV,aAAc,eAChB,CACF,CAAC,EAED,IAAM,EAAc,eACd,EAAa,oBACb,EAAoB,gBAEtB,EAAoB,GAExB,eAAe,GAAyB,CACtC,GAAI,EAAmB,OACvB,EAAoB,GAEpB,IAAM,EAAS,MAAM,OAAO,QAAQ,MAAM,IAAI,CAAC,EAAmB,CAAU,CAAC,EACvE,EAAc,EAAO,GACrB,EAAW,EAAO,GAEpB,GAAe,CAAC,GAAY,CAAC,EAAY,WAAW,KAAK,IAC3D,MAAM,OAAO,QAAQ,MAAM,IAAI,EAAG,GAAa,CAAY,CAAC,EAC5D,MAAM,OAAO,QAAQ,MAAM,OAAO,CAAiB,EAEvD,CAMA,eAAsB,GAAe,CAEnC,IAAM,GAAe,MADA,OAAO,QAAQ,MAAM,IAAI,CAAC,CAAW,CAAC,EAAA,CAC/B,IAAgB,CAAC,EAC7C,MAAO,CAAC,GAAG,EAAgB,GAAG,CAAY,CAC5C,CAMA,eAAsB,GAAkB,CAEtC,OAAO,MADc,OAAO,QAAQ,MAAM,IAAI,CAAC,CAAW,CAAC,EAAA,CAC7C,IAAgB,CAAC,CACjC,CAOA,eAAsB,EAAS,EAAS,CAGtC,OAFK,IAEE,MADW,EAAa,EAAA,CACpB,KAAK,GAAK,EAAE,KAAO,CAAO,GAFhB,IAGvB,CAOA,eAAsB,EAAY,EAAW,CAC3C,IAAM,EAAS,MAAM,EAAgB,EAC/B,EAAW,CACf,GAAI,EAAgB,EACpB,KAAM,EAAU,MAAQ,EAAE,yBAAyB,EACnD,YAAa,EAAU,aAAe,GACtC,KAAM,EAAU,MAAQ,KACxB,aAAc,EAAU,cAAgB,GACxC,QAAS,EAAU,SAAW,KAC9B,SAAU,EAAU,UAAY,KAChC,qBAAsB,EAAU,sBAAwB,KACxD,UAAW,GACX,iBAAkB,EAAU,mBAAqB,IAAA,IAAY,EAAU,iBACvE,MAAO,EAAU,OAAS,KAC1B,YAAa,EAAU,cAAgB,IAAA,GAAoC,KAAxB,EAAU,YAC7D,KAAM,EAAU,OAAS,IAAA,GAA6B,KAAjB,EAAU,KAC/C,UAAW,IAAI,KAAK,CAAA,CAAE,YAAY,EAClC,UAAW,IAAI,KAAK,CAAA,CAAE,YAAY,CACpC,EAGA,OAFA,EAAO,KAAK,CAAQ,EACpB,MAAM,OAAO,QAAQ,MAAM,IAAI,EAAG,GAAc,CAAO,CAAC,EACjD,CACT,CAQA,eAAsB,EAAY,EAAS,EAAS,CAClD,IAAM,EAAS,MAAM,EAAgB,EAC/B,EAAM,EAAO,UAAU,GAAK,EAAE,KAAO,CAAO,EAWlD,OAVI,IAAQ,GAAW,MAEvB,EAAO,GAAO,CACZ,GAAG,EAAO,GACV,GAAG,EACH,GAAI,EACJ,UAAW,GACX,UAAW,IAAI,KAAK,CAAA,CAAE,YAAY,CACpC,EACA,MAAM,OAAO,QAAQ,MAAM,IAAI,EAAG,GAAc,CAAO,CAAC,EACjD,EAAO,GAChB,CAOA,eAAsB,EAAY,EAAS,CACzC,IAAM,EAAS,MAAM,EAAgB,EAC/B,EAAM,EAAO,UAAU,GAAK,EAAE,KAAO,CAAO,EAYlD,OAXI,IAAQ,KAEZ,EAAO,OAAO,EAAK,CAAC,EACpB,MAAM,OAAO,QAAQ,MAAM,IAAI,EAAG,GAAc,CAAO,CAAC,GAIpD,MADiB,OAAO,QAAQ,MAAM,IAAI,CAAC,CAAU,CAAC,EAAA,CAC/C,KAAgB,GACzB,MAAM,OAAO,QAAQ,MAAM,OAAO,CAAU,EAGvC,GACT,CAMA,eAAsB,GAAmB,CAGvC,OAFA,MAAM,EAAuB,GAEtB,MADc,OAAO,QAAQ,MAAM,IAAI,CAAC,CAAU,CAAC,EAAA,CAC5C,IAAe,IAC/B,CAMA,eAAsB,EAAiB,EAAS,CAC1C,EACF,MAAM,OAAO,QAAQ,MAAM,IAAI,EAAG,GAAa,CAAQ,CAAC,EAExD,MAAM,OAAO,QAAQ,MAAM,OAAO,CAAU,CAEhD"}
+{"version":3,"file":"agent-store-DcAs6YyA.js","names":[],"sources":["../../src/side_panel/agent-store.js"],"sourcesContent":["// side_panel/agent-store.js - Agent 数据持久化\nimport { BUILTIN_AGENTS, generateAgentId } from '../shared/agent-defaults.js';\nimport { t, registerTranslations } from '../shared/i18n.js';\n\nregisterTranslations('zh', {\n  agentStore: {\n    unnamedAgent: '未命名Agent',\n  },\n});\nregisterTranslations('en', {\n  agentStore: {\n    unnamedAgent: 'Unnamed Agent',\n  },\n});\n\nconst STORAGE_KEY = 'customAgents';\nconst ACTIVE_KEY = 'activeAssistantId';\nconst LEGACY_ACTIVE_KEY = 'activeAgentId';\n\nlet migratedActiveKey = false;\n\nasync function migrateLegacyActiveKey() {\n  if (migratedActiveKey) return;\n  migratedActiveKey = true;\n\n  const result = await chrome.storage.local.get([LEGACY_ACTIVE_KEY, ACTIVE_KEY]);\n  const legacyValue = result[LEGACY_ACTIVE_KEY];\n  const newValue = result[ACTIVE_KEY];\n\n  if (legacyValue && !newValue && !legacyValue.startsWith('pa_')) {\n    await chrome.storage.local.set({ [ACTIVE_KEY]: legacyValue });\n    await chrome.storage.local.remove(LEGACY_ACTIVE_KEY);\n  }\n}\n\n/**\n * 获取所有 Agent（内置 + 用户自定义）\n * @returns {Promise<Array>}\n */\nexport async function getAllAgents() {\n  const result = await chrome.storage.local.get([STORAGE_KEY]);\n  const customAgents = result[STORAGE_KEY] || [];\n  return [...BUILTIN_AGENTS, ...customAgents];\n}\n\n/**\n * 仅获取用户自定义的 Agent\n * @returns {Promise<Array>}\n */\nexport async function getCustomAgents() {\n  const result = await chrome.storage.local.get([STORAGE_KEY]);\n  return result[STORAGE_KEY] || [];\n}\n\n/**\n * 根据 ID 获取单个 Agent\n * @param {string} agentId\n * @returns {Promise<Object|null>}\n */\nexport async function getAgent(agentId) {\n  if (!agentId) return null;\n  const all = await getAllAgents();\n  return all.find(a => a.id === agentId) || null;\n}\n\n/**\n * 创建新的自定义 Agent\n * @param {Object} agentData - 不含 id 的 Agent 数据\n * @returns {Promise<Object>}\n */\nexport async function createAgent(agentData) {\n  const agents = await getCustomAgents();\n  const newAgent = {\n    id: generateAgentId(),\n    name: agentData.name || t('agentStore.unnamedAgent'),\n    description: agentData.description || '',\n    icon: agentData.icon || '🤖',\n    systemPrompt: agentData.systemPrompt || '',\n    toolIds: agentData.toolIds || null,  // null = 继承全局\n    skillIds: agentData.skillIds || null, // null = 继承全部启用技能\n    mcpExcludedServerIds: agentData.mcpExcludedServerIds ?? null, // MCP 服务排除列表（服务级 deny-list）\n    isBuiltin: false,\n    allowSubDispatch: agentData.allowSubDispatch !== undefined ? agentData.allowSubDispatch : false,\n    model: agentData.model || null,\n    temperature: agentData.temperature !== undefined ? agentData.temperature : null,\n    topP: agentData.topP !== undefined ? agentData.topP : null,\n    createdAt: new Date().toISOString(),\n    updatedAt: new Date().toISOString(),\n  };\n  agents.push(newAgent);\n  await chrome.storage.local.set({ [STORAGE_KEY]: agents });\n  return newAgent;\n}\n\n/**\n * 更新已有的自定义 Agent\n * @param {string} agentId\n * @param {Object} updates\n * @returns {Promise<Object|null>}\n */\nexport async function updateAgent(agentId, updates) {\n  const agents = await getCustomAgents();\n  const idx = agents.findIndex(a => a.id === agentId);\n  if (idx === -1) return null;\n  \n  agents[idx] = {\n    ...agents[idx],\n    ...updates,\n    id: agentId,  // id 不可变\n    isBuiltin: false,  // 不可变\n    updatedAt: new Date().toISOString(),\n  };\n  await chrome.storage.local.set({ [STORAGE_KEY]: agents });\n  return agents[idx];\n}\n\n/**\n * 删除自定义 Agent\n * @param {string} agentId\n * @returns {Promise<boolean>}\n */\nexport async function deleteAgent(agentId) {\n  const agents = await getCustomAgents();\n  const idx = agents.findIndex(a => a.id === agentId);\n  if (idx === -1) return false;\n  \n  agents.splice(idx, 1);\n  await chrome.storage.local.set({ [STORAGE_KEY]: agents });\n\n  // 如果删除的 Agent 是当前选中的，重置为默认\n  const result = await chrome.storage.local.get([ACTIVE_KEY]);\n  if (result[ACTIVE_KEY] === agentId) {\n    await chrome.storage.local.remove(ACTIVE_KEY);\n  }\n  \n  return true;\n}\n\n/**\n * 获取当前选中的 Agent ID\n * @returns {Promise<string|null>}\n */\nexport async function getActiveAgentId() {\n  await migrateLegacyActiveKey();\n  const result = await chrome.storage.local.get([ACTIVE_KEY]);\n  return result[ACTIVE_KEY] || null;\n}\n\n/**\n * 设置当前选中的 Agent ID\n * @param {string|null} agentId\n */\nexport async function setActiveAgentId(agentId) {\n  if (agentId) {\n    await chrome.storage.local.set({ [ACTIVE_KEY]: agentId });\n  } else {\n    await chrome.storage.local.remove(ACTIVE_KEY);\n  }\n}\n\n/**\n * 从模板创建 Agent\n * @param {Object} template\n * @returns {Promise<Object>}\n */\nexport async function createAgentFromTemplate(template) {\n  return createAgent({\n    name: template.name,\n    icon: template.icon,\n    description: template.description,\n    systemPrompt: template.systemPrompt,\n    toolIds: template.toolIds ? [...template.toolIds] : null,\n    allowSubDispatch: template.allowSubDispatch || false,\n  });\n}\n"],"mappings":"+TAIA,EAAqB,KAAM,CACzB,WAAY,CACV,aAAc,UAChB,CACF,CAAC,EACD,EAAqB,KAAM,CACzB,WAAY,CACV,aAAc,eAChB,CACF,CAAC,EAED,IAAM,EAAc,eACd,EAAa,oBACb,EAAoB,gBAEtB,EAAoB,GAExB,eAAe,GAAyB,CACtC,GAAI,EAAmB,OACvB,EAAoB,GAEpB,IAAM,EAAS,MAAM,OAAO,QAAQ,MAAM,IAAI,CAAC,EAAmB,CAAU,CAAC,EACvE,EAAc,EAAO,GACrB,EAAW,EAAO,GAEpB,GAAe,CAAC,GAAY,CAAC,EAAY,WAAW,KAAK,IAC3D,MAAM,OAAO,QAAQ,MAAM,IAAI,EAAG,GAAa,CAAY,CAAC,EAC5D,MAAM,OAAO,QAAQ,MAAM,OAAO,CAAiB,EAEvD,CAMA,eAAsB,GAAe,CAEnC,IAAM,GAAe,MADA,OAAO,QAAQ,MAAM,IAAI,CAAC,CAAW,CAAC,EAAA,CAC/B,IAAgB,CAAC,EAC7C,MAAO,CAAC,GAAG,EAAgB,GAAG,CAAY,CAC5C,CAMA,eAAsB,GAAkB,CAEtC,OAAO,MADc,OAAO,QAAQ,MAAM,IAAI,CAAC,CAAW,CAAC,EAAA,CAC7C,IAAgB,CAAC,CACjC,CAOA,eAAsB,EAAS,EAAS,CAGtC,OAFK,IAEE,MADW,EAAa,EAAA,CACpB,KAAK,GAAK,EAAE,KAAO,CAAO,GAFhB,IAGvB,CAOA,eAAsB,EAAY,EAAW,CAC3C,IAAM,EAAS,MAAM,EAAgB,EAC/B,EAAW,CACf,GAAI,EAAgB,EACpB,KAAM,EAAU,MAAQ,EAAE,yBAAyB,EACnD,YAAa,EAAU,aAAe,GACtC,KAAM,EAAU,MAAQ,KACxB,aAAc,EAAU,cAAgB,GACxC,QAAS,EAAU,SAAW,KAC9B,SAAU,EAAU,UAAY,KAChC,qBAAsB,EAAU,sBAAwB,KACxD,UAAW,GACX,iBAAkB,EAAU,mBAAqB,IAAA,IAAY,EAAU,iBACvE,MAAO,EAAU,OAAS,KAC1B,YAAa,EAAU,cAAgB,IAAA,GAAoC,KAAxB,EAAU,YAC7D,KAAM,EAAU,OAAS,IAAA,GAA6B,KAAjB,EAAU,KAC/C,UAAW,IAAI,KAAK,CAAA,CAAE,YAAY,EAClC,UAAW,IAAI,KAAK,CAAA,CAAE,YAAY,CACpC,EAGA,OAFA,EAAO,KAAK,CAAQ,EACpB,MAAM,OAAO,QAAQ,MAAM,IAAI,EAAG,GAAc,CAAO,CAAC,EACjD,CACT,CAQA,eAAsB,EAAY,EAAS,EAAS,CAClD,IAAM,EAAS,MAAM,EAAgB,EAC/B,EAAM,EAAO,UAAU,GAAK,EAAE,KAAO,CAAO,EAWlD,OAVI,IAAQ,GAAW,MAEvB,EAAO,GAAO,CACZ,GAAG,EAAO,GACV,GAAG,EACH,GAAI,EACJ,UAAW,GACX,UAAW,IAAI,KAAK,CAAA,CAAE,YAAY,CACpC,EACA,MAAM,OAAO,QAAQ,MAAM,IAAI,EAAG,GAAc,CAAO,CAAC,EACjD,EAAO,GAChB,CAOA,eAAsB,EAAY,EAAS,CACzC,IAAM,EAAS,MAAM,EAAgB,EAC/B,EAAM,EAAO,UAAU,GAAK,EAAE,KAAO,CAAO,EAYlD,OAXI,IAAQ,KAEZ,EAAO,OAAO,EAAK,CAAC,EACpB,MAAM,OAAO,QAAQ,MAAM,IAAI,EAAG,GAAc,CAAO,CAAC,GAIpD,MADiB,OAAO,QAAQ,MAAM,IAAI,CAAC,CAAU,CAAC,EAAA,CAC/C,KAAgB,GACzB,MAAM,OAAO,QAAQ,MAAM,OAAO,CAAU,EAGvC,GACT,CAMA,eAAsB,GAAmB,CAGvC,OAFA,MAAM,EAAuB,GAEtB,MADc,OAAO,QAAQ,MAAM,IAAI,CAAC,CAAU,CAAC,EAAA,CAC5C,IAAe,IAC/B,CAMA,eAAsB,EAAiB,EAAS,CAC1C,EACF,MAAM,OAAO,QAAQ,MAAM,IAAI,EAAG,GAAa,CAAQ,CAAC,EAExD,MAAM,OAAO,QAAQ,MAAM,OAAO,CAAU,CAEhD"}