webcode 1.0.0

English

Added workspace file attachments for browser-delivered tool results. attach_file
now supports validated images, PDFs, and UTF-8 TXT files up to 20 MB, preserves
request mapping across result batches, pastes files before result text, and reports
partial delivery failures per request. (#68)
Added risk-aware command approvals. Shell commands are classified as allowed,
confirmable, or blocked, with single-use preflight tokens, separate approval scopes
for shell semantics, and hardened path/option parsing for POSIX and PowerShell. (#67)
Added ChatGPT response-stream tool capture. The browser bridge can extract completed
tool calls directly from response streams, execute network-owned turns with DOM
fallback and deduplication, and show queued, approval, execution, and delivery state
in the shared activity panel. (#66)

Tool activity now also groups DOM-captured calls by assistant message and resets
capture state cleanly across site or configuration lifecycles.
Interactive browser QA now requires explicit user authorization, isolates extension
builds per run, validates local browser and VS Code paths with actionable failures,
and defaults to DeepSeek while preserving supported site selection.

Added agent-led, scenario-driven interactive QA with isolated browser and VS Code
sessions, Playwright inspection, lifecycle checks, run-scoped screenshots, and
agent-owned QA verdicts backed by deterministic evidence. (#65)
Added a deterministic real-task evaluation harness with scenario fixtures, hidden
graders, DeepSeek live-site evaluation, persistent-profile setup controls, traces,
failure attribution, and protocol diagnostics. (#64)
Expanded Extension Host integration coverage and release CI checks so missing compiled
host tests fail explicitly and VS Code test runtimes can be reused across jobs.

新增工作区文件附件能力。attach_file 现在支持经过校验的图片、PDF 和
UTF-8 TXT 文件，单文件最大 20 MB；附件请求可跨结果批次保持映射，文件会
先于结果文本粘贴，并按请求报告部分投递失败。 (#68)
新增基于风险的命令审批。Shell 命令会被分类为允许、需确认或阻止，并通过
一次性预检令牌执行确认；不同 shell 语义使用独立审批范围，同时加强了
POSIX 与 PowerShell 的路径和参数解析。 (#67)
新增 ChatGPT 响应流工具调用捕获。浏览器桥接可以直接从响应流中提取完成的
工具调用，以网络捕获结果为主执行并提供 DOM 回退和去重，同时在统一活动面板
中展示排队、审批、执行和结果投递状态。 (#66)

工具活动面板现在也会按 assistant 消息分组 DOM 捕获的调用，并在站点或配置
生命周期切换时正确重置捕获状态。
交互式浏览器 QA 现在必须经过用户明确授权；每次运行使用隔离的扩展构建，
对本地浏览器和 VS Code 路径给出可操作的失败提示，并默认使用 DeepSeek，
同时保留对受支持站点选择的兼容。

新增由 agent 主导、基于场景的交互式 QA，使用隔离的浏览器与 VS Code 会话，
支持 Playwright 检查、生命周期健康检查、按运行保存截图，并由 agent 给出
QA 结论，确定性证据作为辅助。 (#65)
新增确定性的真实任务评测框架，包含场景夹具、隐藏评分器、DeepSeek 真实站点
评测、持久化 profile 的设置控制、运行 trace、失败归因和协议诊断。 (#64)
扩展 Extension Host 集成覆盖和发布 CI 检查：缺少已编译 Host 测试时会明确失败，
并可在不同任务间复用版本化的 VS Code 测试运行时。