- 
+ 
@@ -31,10 +31,10 @@
## Highlights
- **Built for Local Repositories**: Read, inspect, and edit local codebases with built-in line-by-line diff inspection and auditable checkpoints (Rewind).
-- **Transparent Permission Control**: Choose between read-only analysis, approval-gated runs, and automated Full Access. High-risk actions (file deletion, unlisted commands) always require manual confirmation.
+- **Transparent Permission Control**: Ask / Auto / Full authorization allows, requests approval for, or denies operations according to current permissions, verified effects, and safety policy. Full requires explicit activation in the current process.
- **Ready for Any Model**: Native support for OpenAI, Anthropic Claude, DeepSeek, Kimi, local Ollama models, and custom OpenAI-compatible endpoints with on-the-fly model switching. Also supports quick browser-based ChatGPT account sign-in.
- **Traceable Web Research**: Prioritizes provider-hosted native search, fetches GitHub Issues/PRs, Hacker News, and RSS/Atom feeds, and keeps source snapshots and citations.
-- **Modern Liquid Glass Aesthetic**: Crafted with pure CSS optical refraction (macOS / VisionOS acrylic style), supporting Dark, Light, and Mist frosted themes at a responsive 120fps.
+- **Switchable Interface Themes**: Dark, Light, and Mist themes use CSS blur, translucent backgrounds, gradients, and edge shadows for a glass-like appearance.
---
@@ -110,6 +110,7 @@ Switch between **System, Light, Dark, and Mist** under **Settings → Appearance
| :--- | :--- | :--- |
| **User Guide** | [Usage Guide](docs/usage.md) | Client walkthrough, CLI commands, and typical workflows |
| **Web Research** | [Search & Citations](docs/web-evidence.md) | Search setup, snapshot storage, URL provenance, and diagnostics |
+| **Chat Resources** | [Links, Files & Clipboard](docs/chat-resource-interactions.md) | Link previews, resource context menus, and file/image copy and paste |
| **Code Auditing** | [Workspace Checkpoints](docs/workspace-checkpoints.md) | Line-by-line diff review, snapshots, and safe Rewind |
| **Extensions** | [Extensions & MCP](docs/extensions.md) | Integrating Model Context Protocol (MCP) and custom skills |
| **Code Intelligence** | [LSP Support](docs/code-intelligence.md) | Symbol jump, type inference, and repository index |
diff --git a/README.md b/README.md
index 6c05fa64e..f4ed4e7ab 100644
--- a/README.md
+++ b/README.md
@@ -16,7 +16,7 @@
@@ -31,10 +31,10 @@
## 核心特性
- **深度聚焦本地项目**：直接阅读、分析与修改本地代码库，内置 Diff 逐行审阅与可信撤销（Rewind / Checkpoints）能力。
-- **透明可控的安全策略**：支持只读分析、交互审批（Approval Mode）与全自动（Full Access）权限分级，高危操作（文件删除、非白名单命令）严格阻断并要求人工确认。
+- **透明可控的安全策略**：采用 Ask / Auto / Full 分级授权；操作按当前权限、实际影响和安全策略自动执行、请求审批或拒绝，Full 须在当前进程显式激活。
- **主流模型即开即用**：原生接入 OpenAI、Anthropic Claude、DeepSeek、Kimi、Ollama 本地模型与任何兼容 OpenAI 的第三方 API，支持单会话内即时热切，并支持通过系统浏览器快速完成 ChatGPT 账户授权登录。
- **真实来源的网络检索**：优先调用供应商原生搜索，支持直接抓取 GitHub PR/Issue、Hacker News 与 RSS/Atom 资讯，自动附带快照存证与引用出处。
-- **现代苹果液态玻璃美学**：采用纯 CSS 物理折射模型（macOS / VisionOS 亚克力液态玻璃风格），提供深色、浅色与雾透磨砂质感，兼具 120fps 极速渲染。
+- **可切换的界面主题**：提供深色、浅色与雾透主题，使用 CSS 模糊、半透明背景、渐变和边缘阴影呈现玻璃质感。
---
diff --git a/docs/architecture.md b/docs/architecture.md
index 4c442b8f0..5344b3917 100644
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -157,7 +157,7 @@ The production React bundle is built by Vite but hosted only when Go receives an
## Core Domain
-`Run` is a domain aggregate, not a programming language, operating-system process, or replacement for Go/TypeScript/Rust. Go creates and owns Run lifecycle state; TypeScript may display and control it through Go APIs. Rust never owns a Run. The current analyzer fixture has no Run field or product bridge; a future Go adapter may bind a validated analyzer operation to a Run outside the Rust wire contract.
+`Run` is a domain aggregate, not a programming language, operating-system process, or replacement for Go/TypeScript/Rust. Go creates and owns Run lifecycle state; TypeScript may display and control it through Go APIs. Rust never owns a Run. Go application services bind the fixed embedded digest analyzer to a Run and commit its execution evidence. The Rust wire contract remains free of Run identity and authority. See the [Analyzer README](../analyzers/README.md) for the current execution boundary.
Budget, events, sandbox sessions, and reports are Run-scoped rather than embedded into one large object. Their modules remain independent and persist references by `run_id`; `AgentRunner` coordinates their lifecycle.
@@ -1041,7 +1041,7 @@ internal/store/ SQLite stores and migrations
internal/session/ Compatibility conversation service
internal/tui/ Bubble Tea adapter
internal/httpapi/ Loopback-only read/control API, OpenAPI contract, and bounded Run-event SSE
-internal/analyzer/ Go-owned analyzer protocol, validation, and future bridge boundary
+internal/analyzer/ Go-owned analyzer protocol, validation, and embedded WASI execution boundary
```
Packages split when a concrete implementation establishes different ownership. New tools and integrations reuse these services and add focused behavior tests for their contracts.
@@ -1334,6 +1334,9 @@ records these bounds.
## Go-Owned Analyzer Protocol And Rust Fixture
+The following P10-A/B sections record the initial implementation stages. For the
+current product execution path, see the [Analyzer README](../analyzers/README.md).
+
P10-A1 establishes `internal/analyzer` as a pure Go protocol owner rather than a process
launcher. `analyzer_protocol.v1` accepts one canonical inline Base64 payload plus explicit
limits and four capability declarations that must all be false. Strict decoding rejects
diff --git a/docs/architecture/bounded-project-tests.md b/docs/architecture/bounded-project-tests.md
index 552623f96..70c434581 100644
--- a/docs/architecture/bounded-project-tests.md
+++ b/docs/architecture/bounded-project-tests.md
@@ -1,11 +1,18 @@
# Bounded project tests
-`run_tests` is the root Supervisor's project test entry point in Code Plan and
-Deliver. Review is a workflow state, not an execution phase: the domain has
+`run_tests` is the project test entry point in Code Plan and Deliver for the root
+Supervisor and recorded `specialist_runner` workers with matching WorkerAuthority.
+Review is a workflow state, not an execution phase: the domain has
only `plan` and `deliver` phases.
-The request contains `version: "run-tests.v1"`, a relative directory `target`,
-and `timeout_milliseconds` from 1 through 25000. The host resolves the pinned
+The request contains `version: "run-tests.v1"`. `action: "run"` (the default when
+omitted) takes a relative directory `target` and `timeout_milliseconds` from 1
+through 25000. Longer tests use `action: "start"` with a timeout up to 1800000 ms,
+then `action: "wait"` with the original `job_id`, returned `cursor`, `max_bytes`
+(4–32768), and `wait_milliseconds` (1–5000). Wait observes the existing Job without
+a new target or timeout; repeating start is not a polling operation.
+
+The host resolves the pinned
project test ID or detects a supported project through the existing project
action registry. It chooses the executable, literal arguments, environment,
closed stdin, disabled network intent, output bounds and empty credentials.
@@ -21,10 +28,11 @@ fall back to the host. Dependency installation retains its sandbox-only gate.
The prepared command is pinned in the durable tool authority and regenerated
before approval and dispatch. The existing command authorizer, Run lease,
root attempt, gateway tool budget, command budget, timeout, cancellation and
-Job/output receipts remain authoritative. A test runs as one synchronous Job;
-failed or timed-out Jobs are returned as test evidence. A completed call is
-replayed without starting another Job. Specialist calls retain their existing
-bounded validation tools and child budget; `run_tests` grants no child access.
+Job/output receipts remain authoritative. Each run/start creates one Job; wait
+returns bounded output and status for that Job, retaining its original actor and
+authority. Failed or timed-out Jobs are returned as test evidence. A completed
+call is replayed without starting another Job. Worker execution retains its
+recorded assignment, WorkerAuthority, and child budget checks.
General `command_runtime` remains Deliver-only. Plan Job insertion additionally
requires a pending `run_tests` call with matching actor, command fingerprint,
diff --git a/docs/evidence/markdown-current-state/after.png b/docs/evidence/markdown-current-state/after.png
new file mode 100644
index 0000000000000000000000000000000000000000..728e4730caa90a07b5740ddfbd35a03323e99ded
GIT binary patch
literal 71379
zcmeEuWmsF^^Jk4xpoIbj3KVG3;>9Heic4{~5Q-Lemz1^?EAC$0-8Hzo1q&4S1Pu~M
zwqN<}?)#PhzS#3>?#X?g^O-qw=giERNrC2h>sC3mRl5Sp(40C@Jh
zL_)|&Q4u9N^ul3IV?_r0gOUCWvn*1}$kk6_vt5H{uhFHPNgy1l+<4)Dh|f*abzJG?
zuVZu>+Qy_s(387_5Mj