diff --git a/README.md b/README.md
index d70eac9..abb823f 100644
--- a/README.md
+++ b/README.md
@@ -2,266 +2,161 @@
 
 **English** | [简体中文](README.zh-CN.md)
 
-An open-source web console for creating Agents and running durable conversations
-on Parsar Agent Core or another compatible Agents API Core.
+Agents Core Web is an open-source workspace for creating AI Agents, starting durable
+Sessions, following live work, and managing the resources used by a compatible Agent
+Core. It gives teams a product UI for their own Core deployment without moving
+credentials or execution into the browser.
 
-## What it is
-
-Agents Core Web gives an independently deployed Agent Core a focused browser UI
-and a reusable TypeScript client. The Core remains responsible for authentication,
-persistence, scheduling, and execution; this project does not embed or reimplement it.
+![Agents Core Web Dashboard](docs/images/dashboard.png)
 
 ## What you can do
 
-- Create, view, edit, and delete reusable Agent configurations.
-- Start durable Sessions with optional text input, metadata, and bounded
-  Session-only Agent overrides, then inspect their saved Items.
-- Show Sessions for all Agents or use a server-side root-Agent filter.
-- Optionally create a Codex `self_hosted` Session and follow the connection state
-  of an operator-managed Linux executor.
-- Optionally create the basic Codex `openai_hosted` profile through an
-  operator-qualified managed Runtime, inspect its exact Environment state, and
-  list or explicitly add bounded `/workspace` files.
-- Use Dashboard for the last Agent/Session results successfully traversed to
-  Core's page-chain end marker, and System for live API reachability, pinned
-  contract coverage, Source Files, and ownership boundaries.
-- Follow live progress over SSE and recover persisted output after reconnecting.
-- Cancel active work and return function results or errors.
-- Use the same Web client with Parsar Core or another proven-compatible Core.
-
-## Compatible Agent Cores
-
-| Core or interface | Can Web connect? | Notes |
-| --- | --- | --- |
-| [Parsar Agents API Core at `dadf64a7`](https://github.com/MiniMax-AI-Dev/parsar/tree/dadf64a76bde58255281f3b6c3e939f8b556be09/services/agents-api) | Yes | Primary tested integration and immutable capability baseline |
-| Another Core implementing the tested `/v1/agents/**` HTTP/SSE subset | Yes | It must match the resources and behavior in [protocol coverage](docs/protocol-coverage.md) |
-| OpenAI's hosted Agents API | Not claimed | This project does not promise complete hosted-API compatibility |
-| OpenAI Agents SDK, Responses API, or Parsar daemon WebSocket | No | They are an SDK interface, a model API, and an internal execution interface—not directly connectable Core protocols |
-
-Web speaks a tested subset of the [OpenAI Agents API](https://developers.openai.com/api/docs/guides/agents)
-beta HTTP resource shape:
-JSON requests and authenticated SSE under `/v1/agents/**`, with
-`OpenAI-Beta: agents=v1`. Compatibility means this documented and tested subset,
-not merely accepting the header or sharing similar names.
-
-The capability statements below are audited against immutable Parsar revision
-[`dadf64a7`](https://github.com/MiniMax-AI-Dev/parsar/commit/dadf64a76bde58255281f3b6c3e939f8b556be09),
-not a moving upstream branch.
-
-## Start Web with an existing Core
-
-You need Node.js 22.12+, pnpm 10.30.3, a running compatible Agent Core, and a
-caller bearer issued or configured by that Core's operator. For working chat,
-the Core must also have a configured execution path.
-
-Clone and start the Web:
+- **Operate from one Dashboard** — see loaded Agents, active Sessions, work that needs
+  attention, recent activity, and the two common create flows.
+- **Build reusable Agents** — start from a blank Agent or a practical template, then
+  configure its model, instructions, text behavior, Functions, and HTTP MCP servers.
+- **Run durable conversations** — create a Session, send messages, follow live events,
+  reopen previous work, inspect trace history, and continue after a failed Turn.
+- **Work with tools safely** — review Function calls, submit requested Function results,
+  inspect command and patch activity, and attach write-only MCP credentials through Vaults.
+- **Choose an Environment** — use Core's default execution path, connect a caller-managed
+  self-hosted executor, or use an operator-enabled managed Runtime.
+- **Understand the connection** — view Core reachability and exposed product capabilities,
+  and get actionable local Docker recovery guidance when the backend is not ready.
+
+## Product tour
+
+### Dashboard
+
+Dashboard is the starting point. It summarizes the current Agent and Session results,
+highlights Sessions that need attention, and links directly to Agent creation or a new
+Session.
+
+### Agents
+
+Agents are reusable working profiles. Create one from scratch or begin with a starter
+template for incident response, Slack collaboration, data analysis, GitHub investigation,
+or contract review. Existing Agents open directly in the editor and can start a Session
+from their card.
+
+![Agent library and starter templates](docs/images/agents.png)
+
+### Sessions
+
+Sessions keep the conversation and work history in Core. The workspace combines Session
+navigation, live connection state, conversation output, trace inspection, tool activity,
+and follow-up input without losing the durable record.
+
+![Durable Session conversation](docs/images/sessions.png)
+
+### System and connection status
+
+System shows what the connected Core exposes to this Web build. It separates API access,
+Vault availability, self-hosted presentation, and runtime readiness so a healthy HTTP
+service is not mistaken for a ready model execution path.
+
+![Core connection and capability status](docs/images/system.png)
+
+## Main capabilities
+
+| Area | User experience |
+| --- | --- |
+| Dashboard | Agent and Session overview, attention queue, recent activity, quick actions |
+| Agents | Create, search, inspect, edit, delete, use templates, and start Sessions |
+| Sessions | Durable conversation history, Agent filtering, live events, cancellation, retry and continuation |
+| Trace | Turn history, usage when reported by Core, command output, Function and patch activity |
+| Functions and MCP | Configure supported tools, inspect calls, submit requested results, attach Vault credentials |
+| Vaults | Create project Vaults and manage write-only MCP bearer credentials without reading tokens back |
+| Environments | Default execution, optional self-hosted executor connection, and optional managed Runtime views |
+| Workspace and Files | Inspect supported Environment files and manage project Source Files when enabled by Core |
+| System | Connection status, surfaced capabilities, Core ownership boundaries, and local recovery guidance |
+
+Capabilities appear only when the connected Core and the Web operator configuration expose
+them. Saving an Agent proves that its definition was stored; actual execution still depends
+on the Core's runtime, model provider, credentials, and tool connectivity.
+
+## Quick start
+
+### Requirements
+
+- Node.js 22.12+
+- pnpm 10.30.3
+- a running [compatible Agent Core](#compatible-cores)
+- a caller key issued by that Core's operator
+
+### Start Web
 
 ```bash
 git clone https://github.com/MiniMax-AI-Dev/agents-core-web.git
 cd agents-core-web
 pnpm install
+cp .env.example .env.local
 pnpm dev
 ```
 
-The default local setup expects:
-
-- Agent Core at `http://127.0.0.1:8091`;
-- browser requests through the same-origin `/v1` proxy;
-- the plaintext caller bearer at `~/.parsar/agents-api/web-token`, read only by
-  the local Vite server.
+The default development server opens on `http://127.0.0.1:4173` and proxies browser
+requests to a Core on `http://127.0.0.1:8091`.
 
-Open the loopback URL printed by Vite. Keep the connection base at `/v1`; when
-**Server-managed Core key active** appears, leave the browser token field empty.
-
-To use another server target or private key file, copy `.env.example` to
-`.env.local` and set these optional server-side variables:
+Configure the local proxy in `.env.local`:
 
 ```dotenv
 AGENTS_API_PROXY_TARGET=http://127.0.0.1:8091
 AGENTS_API_PROXY_TOKEN_FILE=/absolute/private/path/to/web-token
 ```
 
-Self-hosted Session creation is a public, non-secret operator opt-in and is hidden
-by default. Enable it only for a reviewed Codex Core deployment whose executor
-registry and externally reachable executor origin are configured:
+Only the local Web server reads the caller-key file. Do not place a plaintext key in a
+`VITE_*` variable, URL, screenshot, Agent, Session metadata, or Git.
 
-```dotenv
-AGENTS_CORE_WEB_SELF_HOSTED_SESSIONS=1
-```
+### First workflow
 
-This flag is read when Vite starts or builds the Web. It exposes the supported
-Session creation form; it does not probe Core capabilities or prove that an
-executor, native runtime, model, or provider is ready. Restart `pnpm dev` after
-changing it. Never place an executor key or any other credential in this variable.
+1. Open **Agents** and choose **Create agent** or a starter template.
+2. Set a name, model, and instructions; add supported tools only when needed.
+3. Select **Start Session**, choose an Environment, and optionally send the first message.
+4. Continue in **Sessions** while live events and durable history update.
+5. Use **Trace**, **Vaults**, **Environment**, or **System** when the workflow needs them.
 
-Basic managed hosted Session creation is a separate default-off presentation
-policy. Enable it only after the Core operator has installed and qualified the
-pinned Codex Runtime image, configured a stable default managed provider, and
-accepted its Docker isolation and model-provider boundary:
+## Local Core and Docker
 
-```dotenv
-AGENTS_CORE_WEB_OPENAI_HOSTED_SESSIONS=1
-```
+When Web cannot reach the local Core, Dashboard opens a connection guide with two paths:
 
-This flag likewise does not discover Core configuration or prove Runtime, native
-harness, model, provider, Function, or tool readiness. The Core may still reject
-creation when no qualified provider is configured.
-
-Keep credentials server-side. A direct Core URL in the connection dialog is only
-for a compatible Core that explicitly allows the Web origin, methods, and headers
-through CORS.
-
-Do not have a Core running yet? Use the immutable
-[current Parsar setup guide](https://github.com/MiniMax-AI-Dev/parsar/blob/dadf64a76bde58255281f3b6c3e939f8b556be09/services/agents-api/README.md#standalone-http-service).
-The repository's [Web connection runbook](docs/core-connection.md) is pinned to the
-same revision and separates ordinary daemon/self-hosted setup from the managed
-Docker-hosted operator profile.
-
-## First use
-
-1. Open **Agents** and create an Agent with a name, instructions, and model ID.
-2. Review, edit, or delete the saved Agent, or open **Start Session**. The default
-   uses no Environment.
-3. Optionally set a title, enter the first text message, or
-   configure the bounded whole-field Agent overrides used only by this Session.
-4. In **Sessions**, choose **All Agents** or one root Agent, select a Session, and
-   continue the conversation.
-5. Follow live Items, cancel active work, or return a requested function result.
-6. Open **System → Source Files** to upload, retrieve, download, or delete one
-   project-owned `user_data` file by its Core ID. Core has no Source Files list, so
-   retain the returned ID; Web does not persist it across page reloads.
-
-**Start Session** accepts an exact non-empty text string or an ordered array of user
-messages containing `input_text` parts only. Message order and grouping are preserved;
-images, attachments, non-user roles, and other content parts are not supported.
-Meaningful input is sent by streaming `POST /v1/agents/sessions` with `stream:true`; Web consumes
-that creation SSE while running durable reconciliation for Session, Item, and
-eligible Environment state and starting the paginated Turn read independently. After
-the POST settles, Web hands live updates off to `GET .../events`. Empty or
-whitespace-only input is omitted and uses the JSON `stream:false` create path for
-`none` and `self_hosted`, so the Session starts idle. Web deliberately uses creation
-SSE for `openai_hosted`, including idle creation, so it can reconcile provisioning
-events before handing off to `GET .../events`. The optional title becomes
-`metadata.title`; Start Session does not expose additional metadata. Additional
-string metadata remains editable from the actions for an existing Session and must
-stay within the pinned Core limits. Never put credentials or secrets there.
-
-Session-only Agent overrides are deliberately finite and whole-field based. Web can
-replace `model`, set or clear `instructions`, replace the plain-text configuration,
-reset saved-only `multi_agent`, `reasoning`, or `service_tier` values to Core
-defaults, and inherit, clear, or fully replace `tools` through the same bounded
-Function/HTTP MCP editor. Untouched fields are omitted. There is no arbitrary Agent
-JSON editor or patch-style partial Tool update. Environment choices are no Environment,
-separately enabled
-`self_hosted`, and separately enabled basic `openai_hosted`; the managed choice is
-blocked when the effective Agent contains MCP because hosted MCP is not qualified.
-
-After a failed create, only an explicit unchanged retry reuses the in-memory
-idempotency key. The stable request fingerprint covers `agent_id`, the finite `agent`
-override when present, `environment`, exact optional `input`, normalized `metadata`,
-`stream`, and sorted derived `vault_ids`; any change to that projected request
-rotates the key. The Sessions root-Agent picker is also server-side: every
-continuation request carries the same `agent_id`, while **All Agents** omits the
-parameter instead of filtering an already loaded page in the browser.
-
-If an HTTP MCP server needs a static bearer, open **Vaults**, create a Vault and a
-write-only Credential for the exact HTTPS endpoint, then select it under
-**Agent → HTTP MCP → Authentication**. When the Session starts, Web attaches the
-owning Vault and never reads the token back. The option is shown only after the
-connected Core successfully exposes the Vault catalog; catalog metadata alone is
-not proof that the MCP server can be reached at runtime.
-
-When the operator opt-in is enabled, **Start Session** also offers **Self-hosted**.
-Its absolute Workspace path is on the executor host, not in the browser, Web server,
-or `parsar-daemon` container. After Core creates the Session, Web can show a
-launcher template built from that Session's Environment ID and executor origin. The
-operator-issued executor credential file stays outside Web, and the launcher itself
-runs on caller-managed Linux executor compute. See
-[Connecting Agent Core](docs/core-connection.md#optional-self-hosted-session-creation)
-for the exact boundary.
-
-The separately opt-in `AGENTS_CORE_WEB_DOCKER_BACKEND_GUIDE=1` profile turns Dashboard
-gateway failures into an explicit recovery entry point. For an existing stack, the
-connection panel shows validated, copyable commands for the configured database, Core
-API, and daemon containers plus the loopback health check. For first-time use, it shows
-the Core image build command and immutable Parsar container/daemon setup links. Parsar
-still requires operator-created database and credential state; Web never runs these
-commands, accesses Docker, or invents secrets. Container names remain non-secret
-operator configuration from `.env.example`.
-
-For the reviewed local loopback stack, an operator can additionally enable the
-default-off `AGENTS_CORE_WEB_DOCKER_GUIDE=1` profile and its required non-secret
-`AGENTS_CORE_WEB_DOCKER_*` settings from `.env.example`. The connection panel then
-offers a copyable Docker command alongside the native launcher. Web still never
-reads the credential file or talks to Docker, and running the command can release
-already queued paid input.
-
-The Source Files surface is independent of Session Environment choice. For a complete
-basic managed Session, Core provisions the `openai_hosted` Runtime automatically;
-Web shows its managed ID, enabled/disabled network policy, empty startup-install
-metadata, and durable/live status. Workspace file controls are independently
-default-off; set `AGENTS_CORE_WEB_ENVIRONMENT_FILES=1` only after qualifying the
-connected Core Files.list and managed Files.create APIs. Managed reads use
-`/workspace`; self-hosted reads use the
-exact Session `workspace_directory`. Web never shows a
-self-hosted launcher or caller connection action for that profile. Inline writes in
-the Session panel and Source-ID copies in System appear only after an exact current
-`openai_hosted` resource read confirms a non-terminal status and empty
-files/plugins/skills metadata. Missing write responses are never replayed. Templates,
-restricted domains, populated startup installs, hosted MCP, readiness discovery, and
-other hosted engines remain unavailable. `self_hosted` Workspace files remain
-read-only. Docker is Core's private managed Runtime adapter, not another public
-Environment discriminator.
-
-The model ID must be supported by the connected execution runtime. Core currently
-has no model-catalog endpoint, so Web suggestions are editable hints rather than
-availability guarantees. Successfully saving an Agent proves configuration storage,
-not that a daemon, model, or provider credential can execute it.
-
-## Relationship to Parsar
-
-```mermaid
-flowchart LR
-  user["User"] --> web["Agents Core Web"]
-  web -->|"HTTP JSON + SSE"| core["Compatible Agent Core"]
-  core --> runtime["Execution runtime / daemon"]
-  runtime --> tools["Models / MCP / tools"]
-```
+- **Already set up** — start the configured database, Core API, and daemon containers,
+  verify `/healthz`, then run **Test connection**.
+- **First time on this computer** — build the Core image and follow the pinned Parsar
+  guides to create the database, caller identity, migrations, API container, and daemon
+  profile.
 
-[Parsar](https://github.com/MiniMax-AI-Dev/parsar) owns its standalone Agents API
-Core, `parsar-daemon`, and native Codex/Claude execution adapters. This repository
-owns only the open Web experience and `@agents-core-web/agents-client`. The browser
-connects to the Core protocol; it never uses the daemon WebSocket as its API URL.
-For `self_hosted`, a separate operator-managed Linux executor connects to Core with
-its own credential and runs commands in its own Workspace; neither Web nor the daemon
-container becomes that Environment.
+Enable the local recovery guide with the non-secret settings documented in
+[`.env.example`](.env.example). Web displays reviewed commands but never receives the
+Docker socket, executes commands, creates credentials, or invents database identities.
 
-See [Architecture](docs/architecture.md) for the full component and trust boundaries.
+For complete setup, see [Connecting Agent Core](docs/core-connection.md). Starting a
+daemon can release queued work and trigger model usage; review pending Sessions first.
 
-## Common problems
+## Compatible Cores
 
-| Symptom | What to check |
+Agents Core Web uses the tested `/v1/agents/**` HTTP and SSE contract documented in
+[Protocol coverage](docs/protocol-coverage.md).
+
+| Core | Status |
 | --- | --- |
-| Web cannot reach Core | Confirm the Core address and `AGENTS_API_PROXY_TARGET`, then restart Vite |
-| `401 invalid_api_key` | The plaintext caller bearer must match the current Core key binding |
-| `503 execution_unavailable` / `Execution is not enabled` | Core rejected execution; inspect its safe error plus runtime and ownership state. A worker, executor, or daemon may be unconfigured or disconnected, or an execution lease may have been lost |
-| Agent saves but its model fails | Use a model ID and provider credential supported by the connected runtime |
-| Self-hosted option is hidden | Set the non-secret `AGENTS_CORE_WEB_SELF_HOSTED_SESSIONS=1` operator flag and restart/rebuild Web only after the connected Codex Core and executor path have been reviewed |
-| Managed hosted option is hidden | Set `AGENTS_CORE_WEB_OPENAI_HOSTED_SESSIONS=1` and restart/rebuild Web only after the pinned Core managed Runtime provider has been qualified |
-| Workspace files are hidden | Set `AGENTS_CORE_WEB_ENVIRONMENT_FILES=1` and restart/rebuild Web only after the connected Core Files.list plus managed Files.create routes and Environment profiles have been qualified |
-
-`/healthz` proves HTTP liveness only, not chat readiness. Check durable Core state and
-the current pinned Parsar guide before retrying an uncertain request; see the
-[connection troubleshooting runbook](docs/core-connection.md#troubleshooting).
+| [Parsar Agents API Core at `dadf64a7`](https://github.com/MiniMax-AI-Dev/parsar/tree/dadf64a76bde58255281f3b6c3e939f8b556be09/services/agents-api) | Primary tested integration |
+| Another Core implementing the documented subset | Compatible after contract qualification |
+| OpenAI hosted Agents API | Complete compatibility is not claimed |
+| OpenAI Agents SDK, Responses API, or Parsar daemon WebSocket | Different interfaces; not direct Core endpoints |
+
+The browser talks to Core. Core owns authentication, durable Agent and Session state,
+scheduling, execution, and runtime resources. The browser never connects directly to a
+daemon or model provider.
 
 ## Documentation
 
-- [Current Parsar Core setup](https://github.com/MiniMax-AI-Dev/parsar/blob/dadf64a76bde58255281f3b6c3e939f8b556be09/services/agents-api/README.md#standalone-http-service) — immutable current upstream guide
-- [Web connection runbook](docs/core-connection.md) — matching `dadf64a7` operator and browser boundary
-- [Protocol coverage](docs/protocol-coverage.md) — exact supported API surface
-- [Architecture](docs/architecture.md) — ownership, runtime, and trust boundaries
-- [Roadmap](docs/roadmap.md) — planned Web and Core integrations
-- [Issue-driven iteration](docs/self-iteration.md) — contributor automation contract
-- [Contributor policy](AGENTS.md) — repository scope, safety, and quality requirements
+- [Connecting Agent Core](docs/core-connection.md) — local, remote, Docker, daemon and Environment setup
+- [Protocol coverage](docs/protocol-coverage.md) — tested resources, events and capability boundaries
+- [Architecture](docs/architecture.md) — components, ownership and trust boundaries
+- [Roadmap](docs/roadmap.md) — planned product and Core integrations
+- [Contributor policy](AGENTS.md) — repository scope, security and quality requirements
+
+The screenshots use isolated local fixture data and do not contain production data or
+prove model-provider readiness.
 
 Agents Core Web is available under the [MIT License](LICENSE).
diff --git a/README.zh-CN.md b/README.zh-CN.md
index 67d7470..34767ad 100644
--- a/README.zh-CN.md
+++ b/README.zh-CN.md
@@ -2,228 +2,151 @@
 
 [English](README.md) | **简体中文**
 
-一个开源 Web 控制台，用于在 Parsar Agent Core 或其他兼容的 Agents API Core
-上创建 Agent 并运行持久化对话。
+Agents Core Web 是一个开源的 AI Agent 工作台，用于创建 Agent、启动可持续的
+Session、跟踪实时执行过程，并管理兼容 Agent Core 提供的资源。团队可以为自己的
+Core 部署提供完整的产品界面，同时让凭据和执行能力始终留在浏览器之外。
 
-## 这是什么
+![Agents Core Web Dashboard](docs/images/dashboard.png)
 
-Agents Core Web 为独立部署的 Agent Core 提供专用的浏览器界面和可复用的
-TypeScript 客户端。鉴权、持久化、调度和执行仍由 Core 负责；本项目不内置或
-重新实现 Agent Core。
+## 可以做什么
 
-## 能做什么
+- **通过 Dashboard 统一管理**：查看 Agent、活跃 Session、需要关注的工作、最近活动
+  和常用创建入口。
+- **创建可复用 Agent**：从空白配置或实用模板开始，设置模型、指令、文本行为、
+  Function 和 HTTP MCP 服务。
+- **运行持久化对话**：创建 Session、发送消息、查看实时事件、重新打开历史工作、
+  检查 Trace，并在 Turn 失败后继续对话。
+- **安全使用工具**：检查 Function 调用、提交被请求的 Function 结果、查看命令和
+  Patch 轨迹，并通过 Vault 绑定只写的 MCP 凭据。
+- **选择执行 Environment**：使用 Core 默认执行链路、连接调用方管理的 self-hosted
+  executor，或使用运维方启用的 managed Runtime。
+- **理解连接状态**：查看 Core 可达性和已暴露能力；本地后端未就绪时获得明确的
+  Docker 恢复步骤。
 
-- 创建、查看、编辑和删除可复用的 Agent 配置。
-- 启动可带初始文本、可选标题和有限 Session 级 Agent 覆盖的持久化 Session，
-  并查看其中保存的 Item。
-- 查看所有 Agent 的 Session，或使用服务端 root-Agent 筛选。
-- 可选创建 Codex `self_hosted` Session，并查看运维方管理的 Linux executor 连接状态。
-- 可选通过运维方已验收的 managed Runtime 创建基础 Codex `openai_hosted`
-  Session，查看精确 Environment 状态，并显式列出或添加有界 `/workspace` 文件。
-- 通过 Dashboard 查看最近一次成功遍历到 Core 分页结束标记的 Agent/Session
-  加载结果，通过 System 查看实时 API 可达性、固定契约范围、Source Files 和所有权边界。
-- 通过 SSE 查看实时进度，并在重连后恢复已持久化的输出。
-- 取消正在执行的任务，并回传函数执行结果或错误。
-- 使用同一个 Web 客户端连接 Parsar Core 或其他经验证兼容的 Core。
+## 产品界面
 
-## 兼容哪些 Agent Core
+### Dashboard
 
-| Core 或接口 | Web 能否连接？ | 说明 |
-| --- | --- | --- |
-| [`dadf64a7` 的 Parsar Agents API Core](https://github.com/MiniMax-AI-Dev/parsar/tree/dadf64a76bde58255281f3b6c3e939f8b556be09/services/agents-api) | 可以 | 主要且经过测试的集成，也是不可变能力基线 |
-| 实现了 `/v1/agents/**` 已测试 HTTP/SSE 子集的其他 Core | 可以 | 必须符合[协议覆盖范围](docs/protocol-coverage.md)记录的资源和行为 |
-| OpenAI 托管的 Agents API | 不作承诺 | 本项目不承诺与托管 API 完全兼容 |
-| OpenAI Agents SDK、Responses API 或 Parsar daemon WebSocket | 不可以 | 它们分别是 SDK 接口、模型 API 和内部执行接口，不是可直接连接的 Core 协议 |
+Dashboard 是默认首页，集中展示当前 Agent 和 Session 结果、需要关注的 Session，
+并可直接进入创建 Agent 或启动 Session 的流程。
 
-Web 使用 [OpenAI Agents API](https://developers.openai.com/api/docs/guides/agents)
-beta HTTP 资源形态中经过测试的子集：
-`/v1/agents/**` 下的 JSON 请求和鉴权 SSE，并携带
-`OpenAI-Beta: agents=v1`。这里的兼容是指已记录、已测试的子集，
-仅仅接受该请求头或名称相似，并不代表兼容。
+### Agents
 
-下文所有能力说明都以不可变 Parsar revision
-[`dadf64a7`](https://github.com/MiniMax-AI-Dev/parsar/commit/dadf64a76bde58255281f3b6c3e939f8b556be09)
-为审计基线，而不是跟随变化的上游分支。
+Agent 是可以反复使用的工作配置。可以从空白 Agent 开始，也可以使用事故响应、
+Slack 协作、数据分析、GitHub 问题调查和合同审阅模板。点击已有 Agent 会直接进入
+编辑，也可以从卡片立即启动 Session。
 
-## 已有 Core 时启动 Web
+![Agent 列表和入门模板](docs/images/agents.png)
 
-你需要 Node.js 22.12+、pnpm 10.30.3、一个正在运行的兼容 Agent Core，
-以及由该 Core 运维方提供或配置的调用方 Bearer 凭据。要正常聊天，Core 还必须
-具备已配置的执行链路。
+### Sessions
 
-克隆并启动 Web：
+Session 的对话和工作历史保存在 Core 中。一个页面同时提供 Session 导航、实时连接
+状态、对话输出、Trace、工具活动和后续输入，不会丢失持久化记录。
+
+![持久化 Session 对话](docs/images/sessions.png)
+
+### System 与连接状态
+
+System 展示当前 Core 对 Web 暴露的能力，并区分 API 访问、Vault、self-hosted 展示
+开关和运行时就绪状态，避免把 HTTP 服务正常误认为模型执行链路已经可用。
+
+![Core 连接和能力状态](docs/images/system.png)
+
+## 主要能力
+
+| 区域 | 用户可以完成的工作 |
+| --- | --- |
+| Dashboard | 查看 Agent/Session 概览、关注队列、最近活动和快捷入口 |
+| Agents | 创建、搜索、查看、编辑、删除、使用模板并启动 Session |
+| Sessions | 持久化对话、按 Agent 筛选、实时事件、取消、重试和继续执行 |
+| Trace | 查看 Turn 历史、Core 报告的 Usage、命令输出、Function 和 Patch 活动 |
+| Functions 与 MCP | 配置支持的工具、查看调用、提交所需结果并绑定 Vault 凭据 |
+| Vaults | 创建项目 Vault，管理只写 MCP Bearer 凭据，Web 不会读回 Token |
+| Environments | 默认执行、可选 self-hosted executor 连接和可选 managed Runtime |
+| Workspace 与 Files | 在 Core 支持时查看 Environment 文件并管理项目 Source Files |
+| System | 查看连接、已暴露能力、Core 所有权边界和本地恢复引导 |
+
+只有连接的 Core 和 Web 运维配置明确暴露的能力才会显示。Agent 保存成功只代表定义
+已经持久化；实际执行仍依赖 Core 的运行时、模型提供商、凭据和工具连接。
+
+## 快速开始
+
+### 环境要求
+
+- Node.js 22.12+
+- pnpm 10.30.3
+- 一个正在运行的[兼容 Agent Core](#兼容的-core)
+- 由 Core 运维方签发的调用方密钥
+
+### 启动 Web
 
 ```bash
 git clone https://github.com/MiniMax-AI-Dev/agents-core-web.git
 cd agents-core-web
 pnpm install
+cp .env.example .env.local
 pnpm dev
 ```
 
-默认本地配置为：
-
-- Agent Core 位于 `http://127.0.0.1:8091`；
-- 浏览器请求通过同源 `/v1` 代理；
-- 明文调用方 Bearer 凭据位于 `~/.parsar/agents-api/web-token`，且只由本地
-  Vite 服务器读取。
-
-打开 Vite 输出的本机回环 URL。连接地址保持为 `/v1`；看到
-**Server-managed Core key active** 时，浏览器 token 输入框应保持为空。
+开发服务器默认使用 `http://127.0.0.1:4173`，并把浏览器请求代理到
+`http://127.0.0.1:8091` 的 Core。
 
-如需使用其他服务地址或私密调用方凭据文件，请将 `.env.example` 复制为
-`.env.local`，并设置以下可选的服务端变量：
+在 `.env.local` 中配置本地代理：
 
 ```dotenv
 AGENTS_API_PROXY_TARGET=http://127.0.0.1:8091
 AGENTS_API_PROXY_TOKEN_FILE=/absolute/private/path/to/web-token
 ```
 
-Self-hosted Session 创建是默认隐藏的非秘密运维开关。只有已经核对 Codex Core、
-executor registry 和 executor origin 的部署才应启用：
+只有本地 Web 服务会读取调用方密钥文件。不要把明文密钥放进 `VITE_*` 变量、URL、
+截图、Agent、Session metadata 或 Git。
 
-```dotenv
-AGENTS_CORE_WEB_SELF_HOSTED_SESSIONS=1
-```
+### 第一个工作流
 
-该开关只暴露已支持的表单，不会探测 Core 能力，也不能证明 executor、原生运行时、
-模型或提供商已就绪。不得在其中放入 executor key 或其他凭据。
+1. 打开 **Agents**，选择 **Create agent** 或一个入门模板。
+2. 设置名称、模型和指令，只在需要时添加支持的工具。
+3. 选择 **Start Session**，选择 Environment，并可同时发送第一条消息。
+4. 在 **Sessions** 中继续工作，查看实时事件和持久化历史。
+5. 工作流需要时再进入 **Trace**、**Vaults**、**Environment** 或 **System**。
 
-基础 managed hosted Session 创建使用另一个默认关闭的展示开关。只有 Core 运维方已
-安装并验收固定 Codex Runtime 镜像、配置稳定的默认 managed provider，并接受 Docker
-隔离和模型提供商边界后才启用：
+## 本地 Core 与 Docker
 
-```dotenv
-AGENTS_CORE_WEB_OPENAI_HOSTED_SESSIONS=1
-```
+Web 无法访问本地 Core 时，Dashboard 会打开连接引导，提供两条路径：
 
-该开关同样不会发现 Core 配置，也不能证明 Runtime、原生 harness、模型、提供商、
-Function 或 Tool 已就绪。Core 未配置合格 provider 时仍会拒绝创建。
-
-修改后重启 `pnpm dev`。凭据应保留在服务端。只有兼容 Core 通过 CORS
-明确允许 Web 的源、方法和请求头时，才能在连接对话框中使用 Core 直连 URL。
-
-还没有运行中的 Core？请使用不可变的
-[当前 Parsar 配置指南](https://github.com/MiniMax-AI-Dev/parsar/blob/dadf64a76bde58255281f3b6c3e939f8b556be09/services/agents-api/README.md#standalone-http-service)。
-仓库内的 [Web 连接手册](docs/core-connection.md)固定在同一 revision，并区分普通
-daemon/self-hosted 配置与 managed Docker-hosted 运维 profile。
-
-## 第一次使用
-
-1. 打开 **Agents**，使用名称、指令（instructions）和 model ID 创建 Agent。
-2. 查看、编辑或删除已保存的 Agent，或打开 **Start Session**；默认不使用
-   Environment。
-3. 可以设置标题、输入第一条文本消息，或配置只作用于该 Session
-   的有限 whole-field Agent 覆盖。
-4. 在 **Sessions** 中选择 **All Agents** 或一个 root Agent，再选择 Session 并继续对话。
-5. 查看实时 Item、取消正在执行的任务，或回传请求的函数结果。
-6. 打开 **System → Source Files**，按 Core 返回的 ID 上传、查询、下载或删除一个
-   project-owned `user_data` 文件。Core 没有 Source Files 列表，Web 也不会在刷新后
-   持久保存该 ID。
-
-**Start Session** 接受精确保留的非空文本字符串，或按原顺序排列、仅包含
-`input_text` part 的 user message 数组；不支持图片、附件、非 user role 或其他
-content part。有效输入通过带 `stream:true` 的流式 `POST /v1/agents/sessions` 发送；Web 一边消费创建 SSE，一边对
-Session、Item 和合格 Environment 的持久状态执行协调，并独立启动 Turn 分页读取。
-POST 结束后，Web 把实时更新交接给 `GET .../events`。对 `none` 和 `self_hosted`，
-空输入或纯空白输入会被省略并走 JSON `stream:false` 创建路径，因此 Session 以 idle
-状态开始。Web 对 `openai_hosted`（包括 idle 创建）固定使用创建 SSE，以便在交接到
-`GET .../events` 前协调 provisioning 事件。可选标题写入 `metadata.title`；
-Start Session 不再暴露额外 metadata。额外字符串 metadata 仍可在已有 Session 的
-Actions 中编辑，并须符合固定 Core 限制；其中绝不能放入凭据或秘密。
-
-Session 级 Agent 覆盖是有限且按 whole-field 生效的集合：Web 可以替换 `model`、
-设置或清空 `instructions`、替换纯文本配置，将 saved-only 的 `multi_agent`、
-`reasoning` 或 `service_tier` 重置为 Core 默认值，并通过同一套受限的
-Function/HTTP MCP 编辑器继承、清空或完整替换 `tools`。未操作的字段不会发送；
-这里没有任意 Agent JSON 编辑器，也不提供 patch 风格的局部 Tool 更新。
-Environment 选项包括无 Environment、单独启用的 `self_hosted` 和单独启用的基础
-`openai_hosted`；effective Agent 含 MCP 时 managed
-选项会被阻止，因为 hosted MCP 尚未验收。
-
-创建失败后，只有显式发起且请求未变化的重试才会复用内存中的幂等 key。稳定
-request fingerprint 完整覆盖 `agent_id`、存在时的有限 `agent` 覆盖、
-`environment`、精确可选 `input`、规范化 `metadata`、`stream` 和排序后的派生
-`vault_ids`；投影请求发生任何变化都会换 key。Sessions 的 root-Agent 选择器同样在
-服务端生效：每个分页请求都携带相同 `agent_id`，而 **All Agents** 会省略该参数，
-不是在浏览器中只筛选已经加载的一页。
-
-HTTP MCP 需要静态 Bearer 时，先在 **Vaults** 中为精确 HTTPS 地址创建 Vault 和
-只写 Credential，再在 **Agent → HTTP MCP → Authentication** 中选择它。创建 Session
-时 Web 会附加所属 Vault，且永不读回 token。只有连接的 Core 成功暴露完整 Vault
-catalog 后才会显示这项能力；catalog 元数据本身不能证明运行时可访问 MCP 服务。
-
-启用运维开关后，**Start Session** 还会提供 **Self-hosted**。Workspace 是 executor
-主机或容器中的绝对路径，不是浏览器、Web 服务或 daemon 容器的目录。Web 只展示
-Core 返回的 Environment ID、executor origin、连接状态和安全 launcher 模板；
-运维方签发的 executor credential 文件始终留在 Web 之外。完整边界见
-[连接 Agent Core](docs/core-connection.md#optional-self-hosted-session-creation)。
-
-可单独启用 `AGENTS_CORE_WEB_DOCKER_BACKEND_GUIDE=1`。Dashboard 遇到网关错误时会明确
-提示 Core 后端未就绪。已有容器时，连接面板会根据 `.env.example` 中经过校验的非秘密
-容器名，展示数据库、Core API、daemon 和 loopback 健康检查的可复制命令；首次使用时，
-则展示 Core 镜像构建命令以及固定 Parsar 版本的容器和 daemon 初始化文档。Parsar 首次
-初始化仍需要运维方创建独立数据库和凭据，Web 不执行命令、不访问 Docker，也不猜测密钥。
-
-对于已核对的本地 loopback 栈，还可以配置 `.env.example` 中默认关闭的
-`AGENTS_CORE_WEB_DOCKER_GUIDE=1` 以及完整的非秘密 `AGENTS_CORE_WEB_DOCKER_*`
-参数。连接面板会在原生 launcher 之外提供可复制的 Docker 命令；Web 仍不会读取
-credential 文件或访问 Docker。若 Session 已有排队输入，运行命令可能立即触发付费调用。
-
-Source Files 与 Session 是否使用 Environment 无关。对于完整的基础 managed Session，
-Core 会自动配置 `openai_hosted` Runtime；Web 展示 managed ID、enabled/disabled 网络
-策略、空 startup-install 元数据、durable/live 状态和显式 `/workspace` 文件列表，绝不
-展示 self-hosted launcher 或调用方连接动作。Session 内联写入和 System 中的 Source-ID
-复制只有在精确查询当前 `openai_hosted` resource、确认非 terminal 状态且
-files/plugins/skills 元数据为空后才显示；写入响应丢失时不会重放。Templates、受限域名、
-非空启动安装、hosted MCP、readiness discovery 和其他 hosted engine 仍不可用。
-`self_hosted` Workspace Files 继续只读。Docker 是 Core 的私有 managed Runtime adapter，
-不是另一个公开 Environment discriminator。
-
-model ID 必须由已连接的执行运行时支持。Core 当前没有模型目录接口，
-因此 Web 建议项只是可编辑提示，不代表模型一定可用。成功保存 Agent 只能证明
-配置已持久化，不能证明 daemon、模型或提供商凭据能够实际执行它。
-
-## 与 Parsar 的关系
-
-```mermaid
-flowchart LR
-  user["User"] --> web["Agents Core Web"]
-  web -->|"HTTP JSON + SSE"| core["Compatible Agent Core"]
-  core --> runtime["Execution runtime / daemon"]
-  runtime --> tools["Models / MCP / tools"]
-```
+- **这台电脑已经配置过**：启动已配置的数据库、Core API 和 daemon 容器，验证
+  `/healthz`，然后运行 **Test connection**。
+- **这台电脑第一次使用**：构建 Core 镜像，并按照固定 Parsar 版本文档创建数据库、
+  调用方身份、迁移、API 容器和 daemon profile。
+
+可使用 [`.env.example`](.env.example) 中的非秘密配置启用本地恢复引导。Web 只展示
+经过校验的命令，不会获得 Docker socket、执行命令、创建凭据或猜测数据库身份。
 
-[Parsar](https://github.com/MiniMax-AI-Dev/parsar) 负责其独立的 Agents API Core、
-`parsar-daemon` 以及原生 Codex/Claude 执行适配器。本仓库只负责开源 Web 体验和
-`@agents-core-web/agents-client`。浏览器连接的是 Core 协议，绝不会把 daemon
-WebSocket 当作 API URL。
+完整步骤参见[连接 Agent Core](docs/core-connection.md)。启动 daemon 可能释放已排队工作
+并产生模型用量，操作前应先检查待处理 Session。
 
-完整的组件和信任边界参见[架构说明](docs/architecture.md)。
+## 兼容的 Core
 
-## 常见问题
+Agents Core Web 使用[协议覆盖范围](docs/protocol-coverage.md)中经过测试的
+`/v1/agents/**` HTTP 和 SSE 契约。
 
-| 现象 | 检查内容 |
+| Core | 状态 |
 | --- | --- |
-| Web 无法访问 Core | 确认 Core 地址和 `AGENTS_API_PROXY_TARGET`，然后重启 Vite |
-| `401 invalid_api_key` | 明文调用方 Bearer 凭据必须与 Core 当前的密钥绑定匹配 |
-| `503 execution_unavailable` / `Execution is not enabled` | Core 拒绝执行；请检查其安全错误、运行时和 ownership 状态。worker、executor 或 daemon 可能未配置或已断连，也可能丢失了执行 lease |
-| Agent 保存成功但模型运行失败 | 使用已连接运行时支持的 model ID 和提供商凭据 |
-| Self-hosted 选项未显示 | 只有核对兼容 Codex Core 与 executor 链路后，设置非秘密 `AGENTS_CORE_WEB_SELF_HOSTED_SESSIONS=1` 并重启或重建 Web |
-| Managed hosted 选项未显示 | 只有固定 Core 的 managed Runtime provider 已验收后，设置 `AGENTS_CORE_WEB_OPENAI_HOSTED_SESSIONS=1` 并重启或重建 Web |
-| Workspace files 未显示 | 只有已验收当前 Core 的 Files.list、managed Files.create 和 Environment profile 后，设置 `AGENTS_CORE_WEB_ENVIRONMENT_FILES=1` 并重启或重建 Web；self-hosted 使用 Session 返回的实际 `workspace_directory` |
-
-`/healthz` 只能证明 HTTP 存活，不能证明聊天已就绪。重试结果不确定的请求前，
-请先核对 Core 持久状态和当前固定版本的 Parsar 指南；参见
-[连接故障排查手册](docs/core-connection.md#troubleshooting)。
+| [`dadf64a7` 的 Parsar Agents API Core](https://github.com/MiniMax-AI-Dev/parsar/tree/dadf64a76bde58255281f3b6c3e939f8b556be09/services/agents-api) | 主要且经过测试的集成 |
+| 实现了已记录子集的其他 Core | 完成契约验收后兼容 |
+| OpenAI 托管 Agents API | 不承诺完全兼容 |
+| OpenAI Agents SDK、Responses API 或 Parsar daemon WebSocket | 属于不同接口，不能作为 Core 地址直接连接 |
+
+浏览器只连接 Core。鉴权、持久化 Agent/Session 状态、调度、执行和运行时资源都由 Core
+负责；浏览器不会直接连接 daemon 或模型提供商。
 
 ## 文档入口
 
-- [当前 Parsar Core 配置](https://github.com/MiniMax-AI-Dev/parsar/blob/dadf64a76bde58255281f3b6c3e939f8b556be09/services/agents-api/README.md#standalone-http-service) — 不可变的当前上游指南
-- [Web 连接手册](docs/core-connection.md) — 对齐 `dadf64a7` 的运维和浏览器边界
-- [协议覆盖范围](docs/protocol-coverage.md) — 准确的已支持 API 范围
-- [架构说明](docs/architecture.md) — 所有权、运行时和信任边界
-- [路线图](docs/roadmap.md) — 计划中的 Web 和 Core 集成
-- [Issue 驱动的自迭代](docs/self-iteration.md) — 贡献者自动化约定
+- [连接 Agent Core](docs/core-connection.md) — 本地、远程、Docker、daemon 和 Environment 配置
+- [协议覆盖范围](docs/protocol-coverage.md) — 已测试资源、事件和能力边界
+- [架构说明](docs/architecture.md) — 组件、所有权和信任边界
+- [路线图](docs/roadmap.md) — 计划中的产品和 Core 集成
 - [贡献者规范](AGENTS.md) — 仓库范围、安全和质量要求
 
+README 截图使用隔离的本地 fixture 数据，不包含生产数据，也不能证明模型提供商已经就绪。
+
 Agents Core Web 采用 [MIT License](LICENSE)。
diff --git a/docs/images/agents.png b/docs/images/agents.png
new file mode 100644
index 0000000..fdffb77
Binary files /dev/null and b/docs/images/agents.png differ
diff --git a/docs/images/dashboard.png b/docs/images/dashboard.png
new file mode 100644
index 0000000..2f1fa21
Binary files /dev/null and b/docs/images/dashboard.png differ
diff --git a/docs/images/sessions.png b/docs/images/sessions.png
new file mode 100644
index 0000000..7edc88c
Binary files /dev/null and b/docs/images/sessions.png differ
diff --git a/docs/images/system.png b/docs/images/system.png
new file mode 100644
index 0000000..7d1fe40
Binary files /dev/null and b/docs/images/system.png differ