feat: 增加记忆后端状态聚合接口

FORK_NOTES.md

@@ -6,6 +6,7 @@ This document explains the fork-specific changes on `main` that diverge from ups
 
 | ID | Target file | What it does | Why we need it | Upstream status |
 |---|---|---|---|---|
+| **P-055** | `agent/memory_provider.py`, `hermes_cli/web_server.py`, `tools/memory_tool.py`, `plugins/memory/openviking/__init__.py`, `plugins/memory/hindsight/__init__.py`, focused tests | Adds an optional read-only runtime-status hook and `GET /api/memory/providers/{name}/status`; OpenViking and Hindsight aggregate health, readiness, component/model/queue/memory or bank/runtime metrics without returning secrets. Provider config saves accept `activate:false`, while omitted `activate` keeps the existing save-and-activate contract. Direct profile-scoped OpenViking endpoints now participate in connection resolution and availability checks. Built-in `MEMORY.md` configuration now publishes and enforces a 1–8000 character range, with the existing 2200-character default. | The managed Desktop runtime has a profile-isolated `HERMES_HOME`; its memory UI could only activate providers and never persisted connection fields, so OpenViking remained `needs_config` and neither backend had in-app observability. Its built-in memory editor also needs one bounded config contract shared with Core. | Generic provider observability/configuration surface; suitable for upstream after UI review. |
 | **P-054** | `model_tools.py`, `tools/environments/base.py`, `tools/environments/local.py`, `tools/environments/modal_utils.py`, `tools/terminal_tool.py`, `tui_gateway/cli_delegation.py`, `tui_gateway/server.py`, Claude Code/Codex skills and focused tests | Extends P-047 to foreground terminal delegations: the agent dispatcher preserves `tool_call_id` through registry dispatch, pipe-backed environments expose a fail-open live-output callback keyed by that id, and the gateway coalesces foreground and background chunks into `delegation.cli.output` at ≤2 Hz. Plain-text output becomes redacted `raw` progress events; completion aggregates model/session/workdir and Token usage from Claude/Codex JSON, with a defensive parser for Codex's human-readable header and `tokens used` footer. Terminal results now include the shell's resolved final cwd, dynamic shell expressions are no longer reported as a literal `$`, and bundled skills request stream-json/`--json` in foreground mode. | A foreground Claude Code/Codex call previously emitted only started/completed, leaving the Desktop spinner silent for minutes; if the completed event was missed the row stayed running forever, while cwd and Token fields remained empty (notably when an older Codex skill omitted `--json`). | Generic terminal/gateway observability and suitable for upstream; event names and classifier remain CN Desktop-driven. Related: P-047. |
 | **P-053** | `scripts/update_thirdparty.py`, `tests/scripts/test_update_thirdparty.py` | Adds a maintenance script that checks GitHub releases for **ripgrep** and **rtk**, updates *every* pinned occurrence across `scripts/install.sh`, `scripts/install.ps1`, `tools/rtk_provision.py`, and `.github/workflows/tests.yml` (which pins ripgrep twice — the `test` and `e2e` jobs), recomputes the CI SHA256 for ripgrep, and supports opt-in Chinese mirror fallback via `--china-mirror` / `--mirror` / `HERMES_THIRDPARTY_MIRROR`. | Hermes-CN pins these external binaries in multiple file formats (Bash, PowerShell, Python, YAML), and direct GitHub access is often slow or blocked in mainland China; a single updater keeps versions consistent and provides mirror fallback. | CN-specific maintenance tool; could be upstreamed after introducing `_ripgrep_common.py` / `_rtk_common.py` |
 | **P-025** | `hermes_cli/web_server.py` | `/api/providers/oauth` now (1) serves from a 20s per-profile in-process TTL cache, (2) runs each provider's status check concurrently via `asyncio.to_thread` (OFF the FastAPI event loop) instead of serially inline, and (3) busts the cache on every connect/disconnect (disconnect clear paths, PKCE submit, device-code/loopback poll→`approved`). Adds a `refresh=true` escape hatch. | The desktop Models page enumerated every OAuth provider's status serially on every open AND every window refocus; some checks touch the network/subprocess, and because the handler is `async` they blocked the event loop that also serves the chat gateway WebSocket — so 模型页 took seconds to open and could stutter live chat. | Should be upstreamed (generic responsiveness fix) |

FORK_NOTES.zh-CN.md

@@ -8,6 +8,7 @@
 
 | ID | 目标文件 | 做了什么 | 为什么需要 | 上游状态 |
 |---|---|---|---|---|
+| **P-055** | `agent/memory_provider.py`、`hermes_cli/web_server.py`、`tools/memory_tool.py`、`plugins/memory/openviking/__init__.py`、`plugins/memory/hindsight/__init__.py`、聚焦测试 | 新增可选只读运行状态 hook 与 `GET /api/memory/providers/{name}/status`；OpenViking/Hindsight 分别聚合健康、就绪、组件、模型、队列、记忆或 bank/runtime 指标，且不返回秘密。Provider 配置保存支持 `activate:false`，省略时仍保持原有“保存并启用”兼容行为。直接写入档案 `memory.openviking` 的 endpoint 也会参与连接解析与可用性判断。内置 `MEMORY.md` 配置继续默认 2200 字符，并统一发布和执行 1–8000 字符范围。 | Desktop managed runtime 使用隔离的 `HERMES_HOME`；旧记忆页只会激活、不会保存连接字段，导致 OpenViking 一直是 `needs_config`，两个后端也都无法在 Hermes 内监控；内置记忆编辑器也需要与 Core 共享同一套容量边界。 | 通用 provider 可观测与配置接口，待 UI 评审后适合上游。 |
 | **P-054** | `model_tools.py`、`tools/environments/base.py`、`tools/environments/local.py`、`tools/environments/modal_utils.py`、`tools/terminal_tool.py`、`tui_gateway/cli_delegation.py`、`tui_gateway/server.py`、Claude Code/Codex 技能及聚焦测试 | 在 P-047 上补齐前台 terminal 委派：agent 调度层把 `tool_call_id` 透传到工具注册表，pipe 型环境再按该 id 提供失败开放的实时输出回调；gateway 将前后台 chunk 都以 ≤2 Hz 合并为 `delegation.cli.output`。普通文本输出归一化为脱敏 `raw` 进度事件；终态从 Claude/Codex JSON 汇总 model/session/workdir 与 Token，并兼容解析 Codex 人类可读头部和 `tokens used` 尾部。terminal 结果补充 shell 执行后的真实 cwd，动态 shell 表达式不再被错误展示为字面量 `$`，内置技能在前台也优先请求 stream-json/`--json`。 | 前台 Claude Code/Codex 过去只有 started/completed，Desktop 会静默转圈数分钟；completed 事件一旦丢失就永久显示运行中，而且旧 Codex 技能未加 `--json` 时目录与 Token 始终为空。 | 通用 terminal/gateway 可观测性可上游；事件名与分类器仍由 CN Desktop 驱动。关联 P-047。 |
 | **P-053** | `scripts/update_thirdparty.py`、`tests/scripts/test_update_thirdparty.py` | 新增维护脚本，检查 **ripgrep** 与 **rtk** 的 GitHub 最新 release，自动同步 `scripts/install.sh`、`scripts/install.ps1`、`tools/rtk_provision.py`、`.github/workflows/tests.yml` 中**每一处**版本钉子（tests.yml 里 `test` 与 `e2e` 两个 job 各钉一次 ripgrep），并重新计算 ripgrep 在 CI 工作流里的 SHA256；支持 `--china-mirror` / `--mirror` / `HERMES_THIRDPARTY_MIRROR` 显式开启的国内镜像回退。 | Hermes-CN 把这些外部二进制版本钉在多种文件格式（Bash、PowerShell、Python、YAML）中，且国内访问 GitHub 经常慢或不通；统一 updater 可保持多文件版本一致，并提供镜像回退。 | CN 专属维护工具；待上游引入 `_ripgrep_common.py` / `_rtk_common.py` 后可考虑 upstream |
 | **P-025** | `hermes_cli/web_server.py` | `/api/providers/oauth` 现在：(1) 命中 20s 的按 profile 进程内 TTL 缓存；(2) 用 `asyncio.to_thread` 并发跑各 provider 的状态检查（移出 FastAPI 事件循环），不再串行内联；(3) 在每次连接/断开时失效缓存（断开的两条清理路径、PKCE submit、设备码/loopback 轮询到 `approved`）。另加 `refresh=true` 逃生阀。 | 桌面端模型页每次打开、以及每次窗口重新聚焦都会串行枚举所有 OAuth provider 的状态；部分检查会联网/起子进程，而该 handler 是 `async`，于是阻塞了同时服务聊天网关 WebSocket 的事件循环——模型页要等好几秒，还会拖累实时会话。 | 建议上游（通用响应性修复） |

agent/memory_provider.py

@@ -28,6 +28,7 @@
   on_pre_compress(messages) -> str       — extract before context compression
   on_memory_write(action, target, content, metadata=None) — mirror built-in memory writes
   on_delegation(task, result, **kwargs)  — parent-side observation of subagent work
+  get_runtime_status() -> dict | None    — read-only backend health and metrics
   backup_paths() -> list[str]            — extra on-disk paths to include in `hermes backup`
 """
 
@@ -260,6 +261,16 @@ def get_config_schema(self) -> List[Dict[str, Any]]:
         """
         return []
 
+    def get_runtime_status(self) -> Optional[Dict[str, Any]]:
+        """Return a read-only backend status snapshot for the dashboard.
+
+        Providers that support monitoring return common connection fields plus
+        a provider-specific ``details`` payload. The dashboard executes this
+        hook off the event loop. Returning ``None`` keeps existing providers
+        compatible and explicitly marks runtime monitoring as unsupported.
+        """
+        return None
+
     def save_config(self, values: Dict[str, Any], hermes_home: str) -> None:
         """Write non-secret config to the provider's native location.
 

agent/models_dev.py

@@ -568,7 +568,12 @@ class ModelCapabilities:
 
     supports_tools: bool = True
     supports_vision: bool = False
+    supports_pdf: bool = False
+    supports_audio: bool = False
+    supports_video: bool = False
     supports_reasoning: bool = False
+    supports_reasoning_control: bool = False
+    open_weights: bool = False
     context_window: int = 200000
     max_output_tokens: int = 8192
     model_family: str = ""
@@ -624,12 +629,14 @@ def get_model_capabilities(
     cache/snapshot only (non-blocking) for hot paths like ``/api/model/info``.
 
     Extracts from model entry fields:
-      - reasoning  (bool)  → supports_reasoning
-      - tool_call  (bool)  → supports_tools
-      - attachment (bool)  → supports_vision
-      - limit.context (int) → context_window
-      - limit.output  (int) → max_output_tokens
-      - family     (str)   → model_family
+      - reasoning          → supports_reasoning
+      - reasoning_options  → supports_reasoning_control
+      - tool_call          → supports_tools
+      - modalities.input   → vision/PDF/audio/video support
+      - open_weights       → open_weights
+      - limit.context      → context_window
+      - limit.output       → max_output_tokens
+      - family             → model_family
     """
     models = _get_provider_models(provider, allow_network=allow_network)
     if models is None:
@@ -653,7 +660,16 @@ def get_model_capabilities(
         supports_vision = "image" in input_mods
     else:
         supports_vision = bool(entry.get("attachment", False))
+    input_modality_set = set(input_mods) if isinstance(input_mods, list) else set()
+    supports_pdf = "pdf" in input_modality_set
+    supports_audio = "audio" in input_modality_set
+    supports_video = "video" in input_modality_set
     supports_reasoning = bool(entry.get("reasoning", False))
+    reasoning_options = entry.get("reasoning_options")
+    supports_reasoning_control = (
+        isinstance(reasoning_options, list) and len(reasoning_options) > 0
+    )
+    open_weights = bool(entry.get("open_weights", False))
 
     # Extract limits
     limit = entry.get("limit", {})
@@ -671,7 +687,12 @@ def get_model_capabilities(
     return ModelCapabilities(
         supports_tools=supports_tools,
         supports_vision=supports_vision,
+        supports_pdf=supports_pdf,
+        supports_audio=supports_audio,
+        supports_video=supports_video,
         supports_reasoning=supports_reasoning,
+        supports_reasoning_control=supports_reasoning_control,
+        open_weights=open_weights,
         context_window=context_window,
         max_output_tokens=max_output_tokens,
         model_family=model_family,

hermes_cli/inventory.py

@@ -147,10 +147,10 @@ def build_models_payload(
       show $/Mtok columns and gate paid models on free accounts —
       mirroring the ``hermes model`` CLI picker. Adds network calls
       (pricing fetch + Nous tier check); only set for interactive pickers.
-    - ``capabilities``: add a per-row ``capabilities`` map
-      ``{model: {fast, reasoning}}`` so pickers can gate the model-options
-      controls (fast toggle / reasoning) to what each model actually
-      supports, instead of offering knobs the backend would reject.
+    - ``capabilities``: add a per-row ``capabilities`` map sourced from the
+      bundled models.dev snapshot. Besides the legacy ``fast``/``reasoning``
+      gates, each known model includes tools, vision, reasoning, context-window,
+      output-token, and family metadata for richer desktop picker tags.
     - ``force_fresh_nous_tier``: bypass the short Nous free-tier cache when
       selecting Portal-recommended Nous models and applying tier gating. Keep
       this false for UI picker opens; explicit auth/model flows can opt in
@@ -264,13 +264,15 @@ def build_models_payload(
 
 
 def _apply_capabilities(rows: list[dict]) -> None:
-    """Attach a ``{model: {fast, reasoning}}`` map to each provider row.
+    """Attach models.dev-backed capability metadata to each provider row.
 
     `fast` mirrors ``model_supports_fast_mode`` (the same gate the runtime
     enforces). `reasoning` comes from the models.dev catalog when known and
     defaults to True otherwise — the effort dial is broadly accepted and a
     no-op on models that ignore it, whereas hiding it from a capable-but-
-    uncatalogued model is the worse failure.
+    uncatalogued model is the worse failure. The richer fields are emitted only
+    when models.dev knows the model, allowing older/static desktop catalogs to
+    remain a fallback for custom or newly-added model IDs.
     """
     from hermes_cli.models import model_supports_fast_mode
 
@@ -281,22 +283,43 @@ def _apply_capabilities(rows: list[dict]) -> None:
 
     for row in rows:
         slug = row.get("slug") or ""
-        caps: dict[str, dict[str, bool]] = {}
+        caps: dict[str, dict[str, object]] = {}
 
         for model in row.get("models") or []:
             reasoning = True
+            meta = None
             if get_model_capabilities is not None and slug:
                 try:
                     meta = get_model_capabilities(slug, model)
+                    if meta is None and str(slug).lower().startswith("custom:"):
+                        canonical_slug = str(slug).split(":", 1)[1].strip().lower()
+                        meta = get_model_capabilities(canonical_slug, model)
                     if meta is not None:
                         reasoning = bool(meta.supports_reasoning)
                 except Exception:
                     reasoning = True
 
-            caps[model] = {
+            model_caps: dict[str, object] = {
                 "fast": bool(model_supports_fast_mode(model)),
                 "reasoning": reasoning,
             }
+            if meta is not None:
+                model_caps.update({
+                    "supports_tools": bool(meta.supports_tools),
+                    "supports_vision": bool(meta.supports_vision),
+                    "supports_pdf": bool(meta.supports_pdf),
+                    "supports_audio": bool(meta.supports_audio),
+                    "supports_video": bool(meta.supports_video),
+                    "supports_reasoning": bool(meta.supports_reasoning),
+                    "supports_reasoning_control": bool(
+                        meta.supports_reasoning_control
+                    ),
+                    "open_weights": bool(meta.open_weights),
+                    "context_window": int(meta.context_window),
+                    "max_output_tokens": int(meta.max_output_tokens),
+                    "model_family": str(meta.model_family or ""),
+                })
+            caps[model] = model_caps
 
         row["capabilities"] = caps