v1.2.6-No-Local-OCR

Release Notes — v1.2.6

重点修复

1. 计划收尾总结不再被吞掉

当计划中存在失败或被取消的步骤时，最终总结会被整个跳过并落到普通问答分支。根因是摘要判定要求「计划已完成」，而任一步骤失败即为未完成。

• 摘要判定放宽为「计划已不再推进」（完成 / 取消 / 含失败步骤均需出总结）；
• 新增兜底流程，三个收尾分支都命中不了时补做最终化，并记录告警便于发现遗漏。

2. 文字交付型步骤不再被误判失败

「回归风险清单与手动验证」这类以文字交付为目的的步骤本就不产生文件修改，此前因回复超过 200 字符闸门被判失败、导致整轮标记为失败。现对长回复增加出口：有 Markdown 结构（标题/列表/表格，≥2 条结构性线索）才认作交付物，纯叙述再长仍判失败。

3. 提示词不再刷进过程折叠块

askStarted 会把整段用户消息/移交提示词（最长数 KB）以 INFO 级别广播给 UI，既泄漏内部脚手架文案，又把折叠块撑大淹没真正的过程信息。现改为只写日志文件，日志仅保留首行 + 字符数与行数统计。

4. 过程折叠块摘要口径修正

摘要原先按时间线文本的非空行数统计，把步骤预告、工具返回、模型中间文本都算了进去，数字还随界面语言与日志格式漂移。现按真实工具调用次数计数，与工具调用回调同源。

5. 修复上下文裁剪因子被重复叠加

同一份 usage 会被 Agent 工具循环与 UI 收尾两条链路各上报一次，导致校准因子被 EMA 连乘两次（实测 24ms 内对同一 usage 输出 1.408 → 1.394）。去重已下沉到唯一的数据源内部，并按 prompt_tokens 标识同一次调用。

---

新功能

• 输入框与消息气泡的 @ 提及高亮：@agent、/ 技能等 token 以蓝色渲染，亮/暗主题各有配色；同时修复选区遮挡与约 1px 的字形错位。
• 模型设置统一表格：原「从端点添加模型 / 自定义模型列表 / 模型最大 Token / 视觉模型」四项合并为一张表格，支持端点获取或手填，并新增来源列、禁止删除官方模型。旧配置无需迁移。
• 思考强度新增 low 档位：补齐官方 low | high | max 枚举（此前仅暴露 high/max），默认仍为 high。
• 构建许可收窄：Edit 计划中仅「标题明确要求构建」或「计划最后一步」可调用构建，避免中间步骤反复触发完整解决方案构建。
• 直接探索预算更宽松：Ask 的直接探索预算由 3 次提升到 5 次；超限后仅提示一次并重新放行，不再永久拦截。

界面

• 修复聊天页面覆盖 VS 自动隐藏弹窗的问题（issue #31 和 issue #88）：WebView2 改用视觉托管，参与正常 z-order。
• 修复 Plan→Edit 交接时过程块不折叠。
• 优化消息载荷：步骤提示中的冗余区块大幅精简（实测某步骤的 user 消息由约 39562 字符降至约 60 字符），并移除代码记忆机制。

国际化

• 修复切换语言后 UI 残留旧语言的三处缺陷（联网搜索开关提示、搜索引擎下拉项、模型下拉的自定义端点后缀）。
• 工具结果、Agent 系统指令与审批卡片等模型可见的硬编码中文全部收口到本地化资源。
• 修复 Ask 机械摘要判断在中文环境下恒为 false 的静默失效问题；补齐少数缺失的键。
• 校验：中英文键集均为 2116 且完全一致，无键缺失导致的 [key] 字面量泄漏。

---

验证

• 单元测试：1729 项全部通过（net472，0 失败 / 0 跳过）。
• 新增测试覆盖：校准去重（7 例）、易变上下文持久化（5 例）、@ 提及分词、构建许可（正向/反向成对）、探索预算、收尾判定等。

Release Notes — v1.2.6

Key Fixes

1. Final summaries are no longer swallowed when a plan contains failed steps

When a plan contains failed or cancelled steps, the final summary was skipped entirely and fell through to the regular Q&A branch. The root cause was that the summary check required "plan completed," while any failed step makes it incomplete.

• The summary check was relaxed to "the plan is no longer progressing" (completed / cancelled / contains failed steps all require a summary);
• A fallback path was added: when none of the three finalization branches match, finalization is performed anyway and a warning is logged to help surface the gap.

2. Text-deliverable steps are no longer wrongly marked as failed

Steps such as "regression risk checklist and manual verification" are meant to deliver text and produce no file changes by design. They were previously rejected by the 200-character gate, marking the entire run as failed. Long responses now have an exit: only replies with Markdown structure (headings/lists/tables, at least 2 structural signals) are accepted as deliverables — pure prose remains a failure no matter how long.

3. Prompts no longer flood the process collapse block

askStarted broadcast the entire user message / handoff prompt (up to several KB) to the UI at INFO level, both leaking internal scaffolding text and inflating the collapse block so real process information was buried. It now writes to the log file only, keeping just the first line plus character and line counts.

4. Corrected the counting basis for the process collapse block summary

The summary previously counted non-empty lines of timeline text, which included step previews, tool returns, and intermediate model text — and the number drifted with the UI language and log format. It now counts actual tool call count, sourced from the same origin as the tool call callback.

5. Fixed duplicate accumulation of the context calibration factor

The same usage was reported once by the Agent tool loop and once by the UI finalization path, causing the calibration factor to be multiplied twice by the EMA (measured: two factors emitted for the same usage within 24ms, 1.408 → 1.394). Deduplication now happens inside the single data source, identifying the same call by prompt_tokens.

---

New Features

• @-mention highlighting in the input box and message bubbles: tokens such as @agent and / skills render in blue, with separate colors for light and dark themes; also fixes selection occlusion and roughly 1px glyph misalignment.
• Unified model settings table: the previous four items ("add model from endpoint / custom model list / model max tokens / vision models") are merged into a single table supporting endpoint fetch or manual entry, plus a new Source column and protection against deleting official models. Existing configuration requires no migration.
• New low reasoning effort level: completes the official low | high | max enum (previously only high/max were exposed); the default remains high.
• Narrower build permission: in an Edit plan, only steps whose title explicitly requires building or the plan's final step may invoke a build, preventing intermediate steps from repeatedly triggering a full solution build.
• More lenient direct exploration budget: the Ask direct exploration budget is raised from 3 to 5; once exceeded, it prompts only once and then allows calls again, instead of blocking permanently.

UI

• Fixed the chat page covering Visual Studio's auto-hidden tool windows (issue #31 and issue #88): WebView2 now uses visual hosting and participates in normal z-order.
• Fixed the process block not collapsing on Plan→Edit handoff.
• Optimized message payloads: redundant blocks in step prompts are greatly trimmed (measured: one step's user message dropped from about 39,562 characters to about 60), and the code memory mechanism was removed.

Internationalization

• Fixed three defects where the UI retained the old language after switching (web search toggle tooltip, search engine dropdown items, and the custom endpoint suffix in the model dropdown).
• Hardcoded Chinese visible to the model — tool results, Agent system instructions, approval cards, and more — is now fully routed through localization resources.
• Fixed a silent failure where the Ask mechanical summary check was always false in the Chinese environment; added a few missing keys.
• Verification: both the Chinese and English key sets contain 2116 keys and are fully identical, with no [key] literal leakage from missing keys.

---

Verification

• Unit tests: all 1729 passed (net472, 0 failed / 0 skipped).
• New test coverage: calibration deduplication (7 cases), volatile context persistence (5 cases), @-mention tokenization, build permission (positive/negative pairs), exploration budget, and finalization decisions.

---

What's Changed

• dev → master：Agent 协作（Handoff 链回）与内置工具增强（Git 范围参数 / 符号工具 / 编辑编码） by @zmy15 in #84
• release: 同步 dev 到 master（编码、UI、网页截图、Git 工具与 EditAgent） by @zmy15 in #85
• feat: 会话过程自动折叠与用量统计修正,Agent 行为修复 by @zmy15 in #86
• fix(chat): WebView2 改用视觉托管，修复覆盖 VS 自动隐藏弹窗 by @zmy15 in #87
• release: v1.2.6 — Agent 收尾/构建许可修复、国际化补全、模型设置统一表格、@ 提及高亮 by @zmy15 in #89

Full Changelog: v1.2.5...v1.2.6-No-Local-OCR

v1.2.6

Release Notes — v1.2.6

重点修复

1. 计划收尾总结不再被吞掉

当计划中存在失败或被取消的步骤时，最终总结会被整个跳过并落到普通问答分支。根因是摘要判定要求「计划已完成」，而任一步骤失败即为未完成。

• 摘要判定放宽为「计划已不再推进」（完成 / 取消 / 含失败步骤均需出总结）；
• 新增兜底流程，三个收尾分支都命中不了时补做最终化，并记录告警便于发现遗漏。

2. 文字交付型步骤不再被误判失败

「回归风险清单与手动验证」这类以文字交付为目的的步骤本就不产生文件修改，此前因回复超过 200 字符闸门被判失败、导致整轮标记为失败。现对长回复增加出口：有 Markdown 结构（标题/列表/表格，≥2 条结构性线索）才认作交付物，纯叙述再长仍判失败。

3. 提示词不再刷进过程折叠块

askStarted 会把整段用户消息/移交提示词（最长数 KB）以 INFO 级别广播给 UI，既泄漏内部脚手架文案，又把折叠块撑大淹没真正的过程信息。现改为只写日志文件，日志仅保留首行 + 字符数与行数统计。

4. 过程折叠块摘要口径修正

摘要原先按时间线文本的非空行数统计，把步骤预告、工具返回、模型中间文本都算了进去，数字还随界面语言与日志格式漂移。现按真实工具调用次数计数，与工具调用回调同源。

5. 修复上下文裁剪因子被重复叠加

同一份 usage 会被 Agent 工具循环与 UI 收尾两条链路各上报一次，导致校准因子被 EMA 连乘两次（实测 24ms 内对同一 usage 输出 1.408 → 1.394）。去重已下沉到唯一的数据源内部，并按 prompt_tokens 标识同一次调用。

---

新功能

• 输入框与消息气泡的 @ 提及高亮：@agent、/ 技能等 token 以蓝色渲染，亮/暗主题各有配色；同时修复选区遮挡与约 1px 的字形错位。
• 模型设置统一表格：原「从端点添加模型 / 自定义模型列表 / 模型最大 Token / 视觉模型」四项合并为一张表格，支持端点获取或手填，并新增来源列、禁止删除官方模型。旧配置无需迁移。
• 思考强度新增 low 档位：补齐官方 low | high | max 枚举（此前仅暴露 high/max），默认仍为 high。
• 构建许可收窄：Edit 计划中仅「标题明确要求构建」或「计划最后一步」可调用构建，避免中间步骤反复触发完整解决方案构建。
• 直接探索预算更宽松：Ask 的直接探索预算由 3 次提升到 5 次；超限后仅提示一次并重新放行，不再永久拦截。

界面

• 修复聊天页面覆盖 VS 自动隐藏弹窗的问题（issue #31 和 issue #88）：WebView2 改用视觉托管，参与正常 z-order。
• 修复 Plan→Edit 交接时过程块不折叠。
• 优化消息载荷：步骤提示中的冗余区块大幅精简（实测某步骤的 user 消息由约 39562 字符降至约 60 字符），并移除代码记忆机制。

国际化

• 修复切换语言后 UI 残留旧语言的三处缺陷（联网搜索开关提示、搜索引擎下拉项、模型下拉的自定义端点后缀）。
• 工具结果、Agent 系统指令与审批卡片等模型可见的硬编码中文全部收口到本地化资源。
• 修复 Ask 机械摘要判断在中文环境下恒为 false 的静默失效问题；补齐少数缺失的键。
• 校验：中英文键集均为 2116 且完全一致，无键缺失导致的 [key] 字面量泄漏。

---

验证

• 单元测试：1729 项全部通过（net472，0 失败 / 0 跳过）。
• 新增测试覆盖：校准去重（7 例）、易变上下文持久化（5 例）、@ 提及分词、构建许可（正向/反向成对）、探索预算、收尾判定等。

Release Notes — v1.2.6

Key Fixes

1. Final summaries are no longer swallowed when a plan contains failed steps

When a plan contains failed or cancelled steps, the final summary was skipped entirely and fell through to the regular Q&A branch. The root cause was that the summary check required "plan completed," while any failed step makes it incomplete.

• The summary check was relaxed to "the plan is no longer progressing" (completed / cancelled / contains failed steps all require a summary);
• A fallback path was added: when none of the three finalization branches match, finalization is performed anyway and a warning is logged to help surface the gap.

2. Text-deliverable steps are no longer wrongly marked as failed

Steps such as "regression risk checklist and manual verification" are meant to deliver text and produce no file changes by design. They were previously rejected by the 200-character gate, marking the entire run as failed. Long responses now have an exit: only replies with Markdown structure (headings/lists/tables, at least 2 structural signals) are accepted as deliverables — pure prose remains a failure no matter how long.

3. Prompts no longer flood the process collapse block

askStarted broadcast the entire user message / handoff prompt (up to several KB) to the UI at INFO level, both leaking internal scaffolding text and inflating the collapse block so real process information was buried. It now writes to the log file only, keeping just the first line plus character and line counts.

4. Corrected the counting basis for the process collapse block summary

The summary previously counted non-empty lines of timeline text, which included step previews, tool returns, and intermediate model text — and the number drifted with the UI language and log format. It now counts actual tool call count, sourced from the same origin as the tool call callback.

5. Fixed duplicate accumulation of the context calibration factor

The same usage was reported once by the Agent tool loop and once by the UI finalization path, causing the calibration factor to be multiplied twice by the EMA (measured: two factors emitted for the same usage within 24ms, 1.408 → 1.394). Deduplication now happens inside the single data source, identifying the same call by prompt_tokens.

---

New Features

• @-mention highlighting in the input box and message bubbles: tokens such as @agent and / skills render in blue, with separate colors for light and dark themes; also fixes selection occlusion and roughly 1px glyph misalignment.
• Unified model settings table: the previous four items ("add model from endpoint / custom model list / model max tokens / vision models") are merged into a single table supporting endpoint fetch or manual entry, plus a new Source column and protection against deleting official models. Existing configuration requires no migration.
• New low reasoning effort level: completes the official low | high | max enum (previously only high/max were exposed); the default remains high.
• Narrower build permission: in an Edit plan, only steps whose title explicitly requires building or the plan's final step may invoke a build, preventing intermediate steps from repeatedly triggering a full solution build.
• More lenient direct exploration budget: the Ask direct exploration budget is raised from 3 to 5; once exceeded, it prompts only once and then allows calls again, instead of blocking permanently.

UI

• Fixed the chat page covering Visual Studio's auto-hidden tool windows (issue #31 and issue #88): WebView2 now uses visual hosting and participates in normal z-order.
• Fixed the process block not collapsing on Plan→Edit handoff.
• Optimized message payloads: redundant blocks in step prompts are greatly trimmed (measured: one step's user message dropped from about 39,562 characters to about 60), and the code memory mechanism was removed.

Internationalization

• Fixed three defects where the UI retained the old language after switching (web search toggle tooltip, search engine dropdown items, and the custom endpoint suffix in the model dropdown).
• Hardcoded Chinese visible to the model — tool results, Agent system instructions, approval cards, and more — is now fully routed through localization resources.
• Fixed a silent failure where the Ask mechanical summary check was always false in the Chinese environment; added a few missing keys.
• Verification: both the Chinese and English key sets contain 2116 keys and are fully identical, with no [key] literal leakage from missing keys.

---

Verification

• Unit tests: all 1729 passed (net472, 0 failed / 0 skipped).
• New test coverage: calibration deduplication (7 cases), volatile context persistence (5 cases), @-mention tokenization, build permission (positive/negative pairs), exploration budget, and finalization decisions.

---

What's Changed

• dev → master：Agent 协作（Handoff 链回）与内置工具增强（Git 范围参数 / 符号工具 / 编辑编码） by @zmy15 in #84
• release: 同步 dev 到 master（编码、UI、网页截图、Git 工具与 EditAgent） by @zmy15 in #85
• feat: 会话过程自动折叠与用量统计修正,Agent 行为修复 by @zmy15 in #86
• fix(chat): WebView2 改用视觉托管，修复覆盖 VS 自动隐藏弹窗 by @zmy15 in #87
• release: v1.2.6 — Agent 收尾/构建许可修复、国际化补全、模型设置统一表格、@ 提及高亮 by @zmy15 in #89

Full Changelog: v1.2.5...v1.2.6

v1.2.5-No-Local-OCR

修复

• 修复超长会话在切换、重启 Visual Studio 或重新应用主题后，消息区域整体白屏的问题（#82）。
• 修复原因是完整会话页面超过 WebView2 NavigateToString 的 2 MB 限制；长会话现在复用窗口化渲染，不再一次性加载全部历史。
• 会话首次打开时优先显示最近 40 条消息，更早内容可通过“加载更早的消息”继续分批查看。
• 保留完整会话树，现有会话无需删除、迁移数据或清理 WebView2 缓存。

优化

• 降低超长会话首次打开和主题重载时的 WebView2 渲染压力，提高页面恢复稳定性。

Fixes

• Fixed the blank message panel affecting long sessions after switching sessions, restarting Visual Studio, or reloading the theme (#82).
• The root cause was that the fully rendered conversation could exceed WebView2's 2 MB NavigateToString limit. Long sessions now reuse the windowed renderer instead of loading the entire history at once.
• Sessions initially display the latest 40 messages, and older messages can be loaded in batches through Load earlier messages.
• The complete conversation tree is preserved. Existing sessions do not require deletion, migration, or WebView2 cache clearing.

Improvements

• Reduced WebView2 rendering pressure when opening very long sessions or reloading the theme, improving panel recovery stability.

What's Changed

• fix(chat): avoid WebView2 navigation size limit by @zmy15 in #83

Full Changelog: v1.2.4...v1.2.5-No-Local-OCR

v1.2.5

修复

• 修复超长会话在切换、重启 Visual Studio 或重新应用主题后，消息区域整体白屏的问题（#82）。
• 修复原因是完整会话页面超过 WebView2 NavigateToString 的 2 MB 限制；长会话现在复用窗口化渲染，不再一次性加载全部历史。
• 会话首次打开时优先显示最近 40 条消息，更早内容可通过“加载更早的消息”继续分批查看。
• 保留完整会话树，现有会话无需删除、迁移数据或清理 WebView2 缓存。

优化

• 降低超长会话首次打开和主题重载时的 WebView2 渲染压力，提高页面恢复稳定性。

Fixes

• Fixed the blank message panel affecting long sessions after switching sessions, restarting Visual Studio, or reloading the theme (#82).
• The root cause was that the fully rendered conversation could exceed WebView2's 2 MB NavigateToString limit. Long sessions now reuse the windowed renderer instead of loading the entire history at once.
• Sessions initially display the latest 40 messages, and older messages can be loaded in batches through Load earlier messages.
• The complete conversation tree is preserved. Existing sessions do not require deletion, migration, or WebView2 cache clearing.

Improvements

• Reduced WebView2 rendering pressure when opening very long sessions or reloading the theme, improving panel recovery stability.

What's Changed

• fix(chat): avoid WebView2 navigation size limit by @zmy15 in #83

Full Changelog: v1.2.4...v1.2.5

v1.2.4-No-Local-OCR

v1.2.4 更新说明

主要更新

模型与 Token 配置

• Token 预算改为百分比

  • 全局统一配置，默认 90%。
  • 有效预算按 模型最大上下文 Token × 预算百分比 计算。
  • 兼容旧版绝对 Token 配置，例如 900000 自动迁移为 90%。

• 支持按模型配置最大 Token

  • 每个官方模型和自定义端点模型均可独立配置。
  • 支持 1M、128K、1000000 等格式。
  • 未配置的模型默认使用 1,000,000 Token。
  • 切换模型后自动重新计算上下文预算。

• 新增模型 Token 选项编辑器

  • 自动列出官方模型、自定义端点模型和已配置模型。
  • 增加来源标记：官方模型、自定义端点、官方 / 自定义端点、已配置。
  • 可直接点击上下文窗口单元格输入配置。

设置界面优化

• 旧版“模型设置”按 01–11 固定顺序显示，不再依赖运行时排序。
• VS2026 新版设置仅保留以下可编辑项：
  • 选择模型
  • 启用深度思考
  • 推理强度
• API 端点、自定义模型、模型最大 Token、视觉模型在新版设置中改为只读引导，提示前往旧版模型设置页。
• 修复旧版设置页面偶尔空白的问题。
• README 增加 GitHub Star 徽章和 Star 提示。

关于信息

• 旧版和新版设置均新增“关于 DeepSeek for Visual Studio”。
• 展示版本、仓库地址和 Issue 地址。
• 支持直接打开仓库、提交 Issue 和为项目点 Star。
• 版本号已更新为 1.2.4。

计价时段

• 所有周末全天均为空闲计价时段。
• 调休上班的周末仍按全天空闲处理。
• 中国法定节假日及调休放假日全天均为空闲计价时段。

问题修复

• 修复 Ask 模式移交 Edit Agent 后，终端命令审批卡片偶尔无法显示的问题。
• 修复终止审批脚本错误、审批卡片互相覆盖的问题。
• 完善通用权限、终端命令和文件删除审批框的并发处理及内容转义。
• 完成 issue 模板中英文内容补全。

测试

• 完整测试集：1471/1471 通过。
• 覆盖模型 Token 预算、自定义端点、设置顺序、关于信息、周末及法定节假日计价。

v1.2.4 Release Notes

Highlights

Model and Token Configuration

• Token budget now uses a percentage

  • Shared globally across all models, with a default of 90%.
  • Effective budget is calculated as model maximum context tokens × budget percentage.
  • Legacy absolute values are migrated automatically, for example 900000 becomes 90%.

• Per-model maximum token configuration

  • Every official model and custom endpoint model can have its own context limit.
  • Supports formats such as 1M, 128K, and 1000000.
  • Models without an explicit value default to 1,000,000 tokens.
  • Context budgets are recalculated immediately when switching models.

• New model token option editor

  • Lists official models, custom endpoint models, and previously configured models.
  • Shows the model source as:
    • Official
    • Custom endpoint
    • Official / Custom endpoint
    • Configured
  • Allows entering the context window directly in the model list.

Settings Improvements

• Legacy model settings now use a fixed 01–11 order instead of runtime sorting.
• VS2026 Unified Settings keeps only these model options editable:
  • Selected Model
  • Enable Deep Thinking
  • Reasoning Effort
• API endpoint, custom models, model maximum tokens, and vision models are now read-only guides that direct users to the legacy model settings page.
• Fixed the legacy settings page occasionally appearing empty.
• Added a GitHub Star badge and Star prompt to the README.

About Information

• Added “About DeepSeek for Visual Studio” to both legacy and Unified Settings.
• Displays the extension version, repository URL, and issue tracker URL.
• Provides actions to open the repository, open the issue tracker, and Star the project.
• Extension version updated to 1.2.4.

Pricing Periods

• All weekends are off-peak for the entire day.
• Weekends designated as adjusted working days remain off-peak all day.
• Chinese public holidays and adjusted rest days are off-peak for the entire day.

Bug Fixes

• Fixed terminal command approval cards occasionally not appearing after an Ask-to-Edit handoff.
• Fixed malformed terminal approval scripts and overlapping approval cards.
• Improved concurrent handling and content escaping for generic permission, terminal command, and file deletion approvals.
• Completed English translations for the issue templates.

Testing

• Full test suite: 1471/1471 passing.
• Coverage includes model token budgets, custom endpoints, settings ordering, About information, weekends, and Chinese public holidays.

What's Changed

• Fix/79 terminal command approval by @zmy15 in #80
• feat: refine pricing periods and settings experience by @zmy15 in #81

Full Changelog: v1.2.3...v1.2.4-No-Local-OCR

v1.2.4

v1.2.4 更新说明

主要更新

模型与 Token 配置

• Token 预算改为百分比

  • 全局统一配置，默认 90%。
  • 有效预算按 模型最大上下文 Token × 预算百分比 计算。
  • 兼容旧版绝对 Token 配置，例如 900000 自动迁移为 90%。

• 支持按模型配置最大 Token

  • 每个官方模型和自定义端点模型均可独立配置。
  • 支持 1M、128K、1000000 等格式。
  • 未配置的模型默认使用 1,000,000 Token。
  • 切换模型后自动重新计算上下文预算。

• 新增模型 Token 选项编辑器

  • 自动列出官方模型、自定义端点模型和已配置模型。
  • 增加来源标记：官方模型、自定义端点、官方 / 自定义端点、已配置。
  • 可直接点击上下文窗口单元格输入配置。

设置界面优化

• 旧版“模型设置”按 01–11 固定顺序显示，不再依赖运行时排序。
• VS2026 新版设置仅保留以下可编辑项：
  • 选择模型
  • 启用深度思考
  • 推理强度
• API 端点、自定义模型、模型最大 Token、视觉模型在新版设置中改为只读引导，提示前往旧版模型设置页。
• 修复旧版设置页面偶尔空白的问题。
• README 增加 GitHub Star 徽章和 Star 提示。

关于信息

• 旧版和新版设置均新增“关于 DeepSeek for Visual Studio”。
• 展示版本、仓库地址和 Issue 地址。
• 支持直接打开仓库、提交 Issue 和为项目点 Star。
• 版本号已更新为 1.2.4。

计价时段

• 所有周末全天均为空闲计价时段。
• 调休上班的周末仍按全天空闲处理。
• 中国法定节假日及调休放假日全天均为空闲计价时段。

问题修复

• 修复 Ask 模式移交 Edit Agent 后，终端命令审批卡片偶尔无法显示的问题。
• 修复终止审批脚本错误、审批卡片互相覆盖的问题。
• 完善通用权限、终端命令和文件删除审批框的并发处理及内容转义。
• 完成 issue 模板中英文内容补全。

测试

• 完整测试集：1471/1471 通过。
• 覆盖模型 Token 预算、自定义端点、设置顺序、关于信息、周末及法定节假日计价。

v1.2.4 Release Notes

Highlights

Model and Token Configuration

• Token budget now uses a percentage

  • Shared globally across all models, with a default of 90%.
  • Effective budget is calculated as model maximum context tokens × budget percentage.
  • Legacy absolute values are migrated automatically, for example 900000 becomes 90%.

• Per-model maximum token configuration

  • Every official model and custom endpoint model can have its own context limit.
  • Supports formats such as 1M, 128K, and 1000000.
  • Models without an explicit value default to 1,000,000 tokens.
  • Context budgets are recalculated immediately when switching models.

• New model token option editor

  • Lists official models, custom endpoint models, and previously configured models.
  • Shows the model source as:
    • Official
    • Custom endpoint
    • Official / Custom endpoint
    • Configured
  • Allows entering the context window directly in the model list.

Settings Improvements

• Legacy model settings now use a fixed 01–11 order instead of runtime sorting.
• VS2026 Unified Settings keeps only these model options editable:
  • Selected Model
  • Enable Deep Thinking
  • Reasoning Effort
• API endpoint, custom models, model maximum tokens, and vision models are now read-only guides that direct users to the legacy model settings page.
• Fixed the legacy settings page occasionally appearing empty.
• Added a GitHub Star badge and Star prompt to the README.

About Information

• Added “About DeepSeek for Visual Studio” to both legacy and Unified Settings.
• Displays the extension version, repository URL, and issue tracker URL.
• Provides actions to open the repository, open the issue tracker, and Star the project.
• Extension version updated to 1.2.4.

Pricing Periods

• All weekends are off-peak for the entire day.
• Weekends designated as adjusted working days remain off-peak all day.
• Chinese public holidays and adjusted rest days are off-peak for the entire day.

Bug Fixes

• Fixed terminal command approval cards occasionally not appearing after an Ask-to-Edit handoff.
• Fixed malformed terminal approval scripts and overlapping approval cards.
• Improved concurrent handling and content escaping for generic permission, terminal command, and file deletion approvals.
• Completed English translations for the issue templates.

Testing

• Full test suite: 1471/1471 passing.
• Coverage includes model token budgets, custom endpoints, settings ordering, About information, weekends, and Chinese public holidays.

What's Changed

• Fix/79 terminal command approval by @zmy15 in #80
• feat: refine pricing periods and settings experience by @zmy15 in #81

Full Changelog: v1.2.3...v1.2.4

v1.2.3-No-Local-OCR

v1.2.3 版本更新

本次更新进一步明确了 DeepSeek for Visual Studio 的 AI Coding Agent 定位，并持续优化编码、调试和设置体验。

更新亮点

• 强化 Coding Agent 定位：将扩展描述更新为 DeepSeek AI coding agent，整合 AI 对话、代码补全和自主编码工作流，为 Visual Studio 2022 与 2026 提供更完整的智能开发体验。
• 新增 2026 年中国节假日计价提示：显示当前 DeepSeek API 处于高峰或空闲计价时段，并自动识别周末及 2026 年中国法定节假日和调休放假日。
• 完善 Visual Studio 快捷键支持：在 DeepSeek Chat 中支持 F5 启动或继续调试、F7 查看代码、Shift+F7 查看设计器，同时避免误触发聊天页面刷新。
• 修复设置页本地化问题：切换语言后，设置页的分组标题和属性描述现在会同步刷新，不再残留旧语言文本。
• 维护与稳定性更新：同步更新 VSIX、README 和 benchmark 中的版本信息，并优化 CI 构建及合并后分支清理流程。

Release v1.2.3

This release strengthens DeepSeek for Visual Studio’s positioning as an AI coding agent and continues to improve the coding, debugging, and configuration experience.

Highlights

• AI Coding Agent positioning: The extension is now described as an AI coding agent that brings chat, code completion, and autonomous coding workflows together in Visual Studio 2022 and 2026.
• 2026 China holiday pricing awareness: Added a peak/off-peak pricing indicator that recognizes weekends, Chinese public holidays, and adjusted rest days for the 2026 holiday schedule.
• Improved Visual Studio shortcuts: DeepSeek Chat now supports F5 to start or continue debugging, F7 to view code, and Shift+F7 to view the designer, without accidentally refreshing the chat page.
• Settings localization fixes: Category titles and property descriptions in the settings page now refresh correctly after changing languages and no longer retain text from the previous language.
• Maintenance and reliability: Updated version metadata across the VSIX package, README, and benchmark configuration, and improved CI build and merged-branch cleanup workflows.

What's Changed

• fix: 完善聊天窗口快捷键、设置本地化与节假日计价，并优化 CI by @zmy15 in #76

Full Changelog: v1.2.3...v1.2.3-No-Local-OCR

v1.2.3

v1.2.3 版本更新

本次更新进一步明确了 DeepSeek for Visual Studio 的 AI Coding Agent 定位，并持续优化编码、调试和设置体验。

更新亮点

• 强化 Coding Agent 定位：将扩展描述更新为 DeepSeek AI coding agent，整合 AI 对话、代码补全和自主编码工作流，为 Visual Studio 2022 与 2026 提供更完整的智能开发体验。
• 新增 2026 年中国节假日计价提示：显示当前 DeepSeek API 处于高峰或空闲计价时段，并自动识别周末及 2026 年中国法定节假日和调休放假日。
• 完善 Visual Studio 快捷键支持：在 DeepSeek Chat 中支持 F5 启动或继续调试、F7 查看代码、Shift+F7 查看设计器，同时避免误触发聊天页面刷新。
• 修复设置页本地化问题：切换语言后，设置页的分组标题和属性描述现在会同步刷新，不再残留旧语言文本。
• 维护与稳定性更新：同步更新 VSIX、README 和 benchmark 中的版本信息，并优化 CI 构建及合并后分支清理流程。

Release v1.2.3

This release strengthens DeepSeek for Visual Studio’s positioning as an AI coding agent and continues to improve the coding, debugging, and configuration experience.

Highlights

• AI Coding Agent positioning: The extension is now described as an AI coding agent that brings chat, code completion, and autonomous coding workflows together in Visual Studio 2022 and 2026.
• 2026 China holiday pricing awareness: Added a peak/off-peak pricing indicator that recognizes weekends, Chinese public holidays, and adjusted rest days for the 2026 holiday schedule.
• Improved Visual Studio shortcuts: DeepSeek Chat now supports F5 to start or continue debugging, F7 to view code, and Shift+F7 to view the designer, without accidentally refreshing the chat page.
• Settings localization fixes: Category titles and property descriptions in the settings page now refresh correctly after changing languages and no longer retain text from the previous language.
• Maintenance and reliability: Updated version metadata across the VSIX package, README, and benchmark configuration, and improved CI build and merged-branch cleanup workflows.

What's Changed

• fix: 完善聊天窗口快捷键、设置本地化与节假日计价，并优化 CI by @zmy15 in #76

Full Changelog: v1.2.2...v1.2.3

v1.2.2-No-Local-OCR

v1.2.2

重点更新

• 新增引导模式和排队模式，支持任务执行期间追加用户消息。
• 重构 Agent 编辑流程，统一使用原生编辑工具，并移除旧文本编辑格式重试流程。
• read_file 对图片启用视觉直传；视觉模型未明确要求 OCR 时隐藏 OCR 工具。
• 中断回复支持原地重试或编辑，不再自动创建新分支。
• 完善设置默认值、设置迁移、恢复默认、本地化资源和统一设置同步。
• 增加 token 吞吐量、Agent 遥测、Git 上下文和终端任务可观测性。

新增与改进

• 支持 request_handoff 的轻量编辑步骤计划，并显示执行进度。
• 用户提问时注入当前 Git 分支和最新提交状态。
• 支持追加消息的持久化、渲染和会话恢复。
• 长文本粘贴自动转为附件。
• 优化 Plan/Edit 上下文复用、Handoff 校验和编辑工具批量原子回滚。
• 改进编译错误输出、构建结果复用和任务完成通知流程。
• 优化思考流式增量、终端编码、分离进程输出和会话切换行为。
• 分支切换按钮移动到 AI 消息的复制/重试操作行。
• 中英文 README 增加界面截图。

修复

• 修复 apply_patch 已成功执行却被误判为没有编辑的问题。
• 修复切换英文设置界面时的枚举转换异常。
• 修复 run_in_terminal URL、终端编码和 Handoff 描述传递问题。
• 修复 Build 与 Ask 之间可能重复移交或循环的问题。
• 修复会话记忆绑定、删除会话后记忆清理和自动记忆准确性问题。
• 修复会话切换后残留文件变更记录导致的错误回退提示。
• 修复 Windows 10 图标字体兼容性和会话删除图标。
• 修复编译错误被截断为 30 条的问题，统一上限为 500 条并提示遗漏数量。

行为变更

• 旧文本编辑格式重试已移除，编辑必须通过可用工具执行。
• 编辑消息只保留上方内联编辑区，不再同步填充输入框。
• 自动记忆默认仅注入 User 作用域。
• 视觉模型未明确要求 OCR 时，不再暴露 OCR 工具；明确请求 OCR 时仍可使用。

验证

• 完整自动化测试通过：1397 passed, 0 failed
• 构建通过：0 errors

Highlights

• Added guided and queued follow-up message modes for sending additional requests while a task is running.
• Reworked the Agent editing flow around native editing tools and removed the legacy text-edit retry path.
• Added direct visual passthrough for images read through read_file; OCR tools are hidden for vision models unless OCR is explicitly requested.
• Interrupted responses can now be retried or edited in place without creating a new branch.
• Improved settings defaults, migration, restore-default behavior, localization, and unified settings synchronization.
• Added token throughput metrics, Agent telemetry, Git context, and better terminal task observability.

Added and Improved

• Added lightweight editSteps plans for request_handoff, including execution progress.
• Injected the current Git branch and latest commit state into user requests.
• Added persistence, rendering, and session restoration for follow-up messages.
• Long pasted text is automatically stored as an attachment.
• Improved Plan/Edit context reuse, Handoff validation, and atomic rollback for batch edits.
• Improved build error reporting, build-result reuse, and task completion notifications.
• Improved reasoning stream handling, terminal encoding, detached process output, and session switching.
• Moved the branch switcher into the AI message copy/retry action row.
• Added interface screenshots to the English and Chinese READMEs.

Fixed

• Fixed successful apply_patch operations being incorrectly reported as having made no edits.
• Fixed enum conversion failures when switching the settings UI to English.
• Fixed run_in_terminal URL handling, terminal encoding, and incomplete Handoff task descriptions.
• Fixed duplicated Build-to-Ask handoffs and possible Handoff loops.
• Fixed session memory binding, memory cleanup after session deletion, and automatic memory accuracy issues.
• Fixed stale file-change records causing incorrect rollback prompts after switching sessions.
• Fixed Windows 10 icon font compatibility and the session deletion icon.
• Fixed build errors being truncated to 30 entries; the limit is now 500 with an omitted-entry notice.

Behavior Changes

• Legacy text-edit format retries were removed; edits now require the available editing tools.
• Edited messages now use the inline editor only and no longer populate the main input box.
• Automatic memory injection is limited to User scope by default.
• OCR tools are hidden from vision models unless OCR is explicitly requested.

Verification

• Full automated test suite: 1397 passed, 0 failed
• Build: 0 errors

What's Changed

• fix(ui): add Windows 10 icon font fallback(#70 ) by @zmy15 in #72
• fix: 自动记忆准确性 + Agent 移交稳定性修复（含 editSteps 轻量计划） by @zmy15 in #73
• feat(agent): 稳定 Agent 工作流并完善 Git 同步/剪贴板附件 by @zmy15 in #74
• release: DeepSeek for Visual Studio v1.2.2 by @zmy15 in #75

Full Changelog: v1.2.2...v1.2.2-No-Local-OCR

v1.2.2

v1.2.2

重点更新

• 新增引导模式和排队模式，支持任务执行期间追加用户消息。
• 重构 Agent 编辑流程，统一使用原生编辑工具，并移除旧文本编辑格式重试流程。
• read_file 对图片启用视觉直传；视觉模型未明确要求 OCR 时隐藏 OCR 工具。
• 中断回复支持原地重试或编辑，不再自动创建新分支。
• 完善设置默认值、设置迁移、恢复默认、本地化资源和统一设置同步。
• 增加 token 吞吐量、Agent 遥测、Git 上下文和终端任务可观测性。

新增与改进

• 支持 request_handoff 的轻量编辑步骤计划，并显示执行进度。
• 用户提问时注入当前 Git 分支和最新提交状态。
• 支持追加消息的持久化、渲染和会话恢复。
• 长文本粘贴自动转为附件。
• 优化 Plan/Edit 上下文复用、Handoff 校验和编辑工具批量原子回滚。
• 改进编译错误输出、构建结果复用和任务完成通知流程。
• 优化思考流式增量、终端编码、分离进程输出和会话切换行为。
• 分支切换按钮移动到 AI 消息的复制/重试操作行。
• 中英文 README 增加界面截图。

修复

• 修复 apply_patch 已成功执行却被误判为没有编辑的问题。
• 修复切换英文设置界面时的枚举转换异常。
• 修复 run_in_terminal URL、终端编码和 Handoff 描述传递问题。
• 修复 Build 与 Ask 之间可能重复移交或循环的问题。
• 修复会话记忆绑定、删除会话后记忆清理和自动记忆准确性问题。
• 修复会话切换后残留文件变更记录导致的错误回退提示。
• 修复 Windows 10 图标字体兼容性和会话删除图标。
• 修复编译错误被截断为 30 条的问题，统一上限为 500 条并提示遗漏数量。

行为变更

• 旧文本编辑格式重试已移除，编辑必须通过可用工具执行。
• 编辑消息只保留上方内联编辑区，不再同步填充输入框。
• 自动记忆默认仅注入 User 作用域。
• 视觉模型未明确要求 OCR 时，不再暴露 OCR 工具；明确请求 OCR 时仍可使用。

验证

• 完整自动化测试通过：1397 passed, 0 failed
• 构建通过：0 errors

Highlights

• Added guided and queued follow-up message modes for sending additional requests while a task is running.
• Reworked the Agent editing flow around native editing tools and removed the legacy text-edit retry path.
• Added direct visual passthrough for images read through read_file; OCR tools are hidden for vision models unless OCR is explicitly requested.
• Interrupted responses can now be retried or edited in place without creating a new branch.
• Improved settings defaults, migration, restore-default behavior, localization, and unified settings synchronization.
• Added token throughput metrics, Agent telemetry, Git context, and better terminal task observability.

Added and Improved

• Added lightweight editSteps plans for request_handoff, including execution progress.
• Injected the current Git branch and latest commit state into user requests.
• Added persistence, rendering, and session restoration for follow-up messages.
• Long pasted text is automatically stored as an attachment.
• Improved Plan/Edit context reuse, Handoff validation, and atomic rollback for batch edits.
• Improved build error reporting, build-result reuse, and task completion notifications.
• Improved reasoning stream handling, terminal encoding, detached process output, and session switching.
• Moved the branch switcher into the AI message copy/retry action row.
• Added interface screenshots to the English and Chinese READMEs.

Fixed

• Fixed successful apply_patch operations being incorrectly reported as having made no edits.
• Fixed enum conversion failures when switching the settings UI to English.
• Fixed run_in_terminal URL handling, terminal encoding, and incomplete Handoff task descriptions.
• Fixed duplicated Build-to-Ask handoffs and possible Handoff loops.
• Fixed session memory binding, memory cleanup after session deletion, and automatic memory accuracy issues.
• Fixed stale file-change records causing incorrect rollback prompts after switching sessions.
• Fixed Windows 10 icon font compatibility and the session deletion icon.
• Fixed build errors being truncated to 30 entries; the limit is now 500 with an omitted-entry notice.

Behavior Changes

• Legacy text-edit format retries were removed; edits now require the available editing tools.
• Edited messages now use the inline editor only and no longer populate the main input box.
• Automatic memory injection is limited to User scope by default.
• OCR tools are hidden from vision models unless OCR is explicitly requested.

Verification

• Full automated test suite: 1397 passed, 0 failed
• Build: 0 errors

What's Changed

• fix(ui): add Windows 10 icon font fallback by @zmy15 in #72
• fix: 自动记忆准确性 + Agent 移交稳定性修复（含 editSteps 轻量计划） by @zmy15 in #73
• feat(agent): 稳定 Agent 工作流并完善 Git 同步/剪贴板附件 by @zmy15 in #74
• release: DeepSeek for Visual Studio v1.2.2 by @zmy15 in #75

Full Changelog: v1.2.1...v1.2.2