DeepSeekEyes v0.5.2

DeepSeekEyes 0.5.2

• Match the native DeepSeek Harness plugin-card disclosure pattern with a full-width accessible header button and rotating chevron.
• Keep the DeepSeekEyes settings card collapsed by default so the Plugins page remains compact alongside built-in cards.
• Preserve staged edits when collapsing, show an unsaved badge, and retain existing live-save behavior after reopening.

Verification

• npm run check: 92/92 tests, fixture eval schema validity 1, accuracy 1
• macOS live DSH settings page: collapsed by default; expand/collapse and chevron verified
• rollback rehearsal: restored v0.5.1 behavior and clean tree

DeepSeekEyes 0.5.1

DeepSeekEyes 0.5.1

This patch stabilizes native macOS application and window routing in Desktop Computer Use.

• launch works directly without a prior observe or stateId.
• Resolve display names, renamed aliases, bundle IDs and full .app paths through Launch Services.
• Prefer the exact application PID and its focused/main usable window instead of helper processes or tiny auxiliary dialogs.
• Explicit application/title targets override stale capture targets.
• Window screenshots and metadata now retain identical dimensions.
• Accessibility discovery is incrementally bounded by element count and helper time, with semanticStatus.truncated, limitReason and elapsedMs.
• Timeout and routing errors now identify the action, target, failed stage and available candidates.
• Read-only window-ref observation is stateless; state-changing ref actions remain latest-state protected.

Verification

• 92/92 local tests pass; 83.19% line coverage.
• Native ChatGPT launch resolves all four forms: ChatGPT, Codex, com.openai.codex, /Applications/ChatGPT.app.
• Real DSH call { "action": "launch", "application": "ChatGPT" } returns the 2018×1141 ChatGPT main window and matching lossless screenshot without state or helper-timeout errors.
• GitHub CI passed on Ubuntu, macOS and Windows, including native desktop observation.
• npm audit --omit=dev: 0 vulnerabilities.

DeepSeekEyes 0.5.0

• Add window-scoped native observation on Windows and macOS, with screenshot-relative coordinate mapping back to the desktop and explicit failure instead of silently returning the wrong scope.
• Add Windows UI Automation and macOS Accessibility element discovery with stable elementRef identities, semantic role/name/value/bounds/state metadata and configurable observation limits.
• Re-resolve macOS Accessibility actions by a stable semantic fingerprint and duplicate ordinal, using the prior flat index only after identity/bounds verification so a shifted tree fails closed instead of acting on another control.
• Prefer semantic elementRef actions while retaining pixel-coordinate fallback: click/type/scroll now accept elements, and new invoke, set_value and perform_action operations call native accessibility patterns.
• Derive stable windowRef values from native window identity while keeping all refs bound to the latest stateId.
• Bind macOS windows to stable CoreGraphics window numbers and re-resolve the matching Accessibility window after z-order changes, preventing a focus/raise action from capturing another window of the same application.
• Return stateDelta after every observation, including pixel-identity screenshot changes and added/removed/changed windows and elements.
• Add native runtime assertions for window existence/title, element existence/visibility/enabled/focus/value/name and screen changed/unchanged, while retaining explicit model-backed visual assertions.
• Upgrade desktop reports to deepseekeyes.desktop-report.v2; typed text, assigned values and launch arguments remain plaintext-free in persisted events.
• Add GUI settings for semantic desktop observation and element limits, increase the default desktop timeout to 30 seconds, and document the 0.5 screenshot/action/feedback loop.
• Extend unit and native acceptance coverage for window capture, semantic refs/actions, state differences, runtime assertions and macOS Accessibility; Windows helper parsing/native observation remain covered by cross-platform CI.
• Keep Windows desktop discovery valid when no window-scoped semantic target exists by normalizing the UI Automation result to an explicit empty array before counting controls.

DeepSeekEyes v0.4.2

Visual-route compatibility fix

This release fixes the 0.2 → 0.4 compatibility regression seen on working MiniMax and gateway vision routes.

• restores extraction of one unique JSON object for visual evidence and the active pixel probe;
• replaces the full in-prompt Schema dump with a compact shape generated from the same canonical Schema;
• normalizes Qwen 0–1000 xyxy, normalized/pixel xyxy, and pixel xywh boxes before strict Ajv validation;
• records original and normalized coordinates without changing the source image or making a repair-model call;
• resolves an unset DSH_HOME to Harness's ~/.dsh default;
• keeps pure-text turns and normal DeepSeek usage on the unchanged direct path.

Upgrade and restart DSH:

npx -y @dttxorg/deepseekeyes@latest upgrade

DeepSeekEyes 0.4.1

Fixes

• Fix scoped npm client-module registration so DSH resolves the settings client after install.
• Keep dense desktop screenshot evidence bounded while preserving original lossless pixels for targeted rereads.
• Retry explicit visual max-token rejections once with provider-managed output.
• Report VISION_OUTPUT_TRUNCATED and the ordered visual route/error-code chain instead of a generic failover counter.

Verified on Ubuntu, Windows and macOS; 84 tests passed.

DeepSeekEyes 0.4.0

DeepSeekEyes 0.4.0

DeepSeekEyes is now the auditable vision and cross-platform Computer Use runtime for DeepSeek Harness.

P0

• Public scoped package: @dttxorg/deepseekeyes
• One-line install, upgrade, and doctor CLI commands
• One canonical strict JSON Schema for initial and targeted visual evidence
• Ajv validation for every nested OCR, region, object, observation, bbox, confidence, and unknown field

P1

• Ordered visual routes, metadata health TTL, failed-route cooldown, bounded failover, and privacy-reduced attempts audit
• Public screenshot, dense-text, chart, UI, and prompt-injection eval with accuracy, latency, Schema, and Token reporting
• SECURITY, Troubleshooting, architecture, data-retention, and npm provenance documentation
• Native GUI reliability controls and continued zero-overhead pure-text path

Verification

• 79 tests passed; 0 failed
• Ubuntu, macOS, and Windows CI passed
• Browser live acceptance and macOS native Desktop observe passed
• Real isolated DSH install, boot, HTTP 200, original-image native model switch, token accounting, and Desktop Computer Use passed
• npm audit --omit=dev: 0 vulnerabilities

Install after npm publication

npx -y @dttxorg/deepseekeyes@latest install
npx -y @dttxorg/deepseekeyes@latest upgrade
npx -y @dttxorg/deepseekeyes@latest doctor

DeepSeekEyes 0.3.1

DeepSeekEyes 0.3.1

本版本在 Harness 原生 设置 → 插件 → DeepSeekEyes 卡片中加入本插件 Token 消耗统计：

• 区分 Provider 精确返回的视觉、探针、追问用量与插件桥接输入估算；
• 最终回答模型的正常回答用量单独显示，不计入插件额外消耗；
• 纯文字轮次继续走原有直通路径，统计增量为 0；
• 刷新、清零、开关统计使用本机回环 RPC，不调用任何模型；
• 统计默认原子写入 DSH_HOME/deepseekeyes/usage-stats.json，权限 0600；
• 写入失败时继续在内存计数，不中断正常对话，后续写入成功时自动恢复落盘。

同时保留 0.3.0 的 Windows/macOS Desktop Computer Use 与 0.2.0 Browser Computer Use。

DeepSeekEyes 0.3.0

DeepSeekEyes 0.3.0

DeepSeekEyes 现已提供 Windows 与 macOS 原生 Computer Use，并继续在同一个 DeepSeek Harness 对话内使用视觉模型读取每次操作后的无损桌面截图。

新增

• 独立 computer 工具：观察、鼠标、键盘、滚动、应用启动、窗口控制、等待、视觉断言与测试报告。
• Windows 原生驱动：PowerShell、user32、SendInput、DPI 感知、Unicode 输入与虚拟桌面坐标。
• macOS 原生驱动：JXA、CoreGraphics、System Events、Unicode 输入与系统截图。
• 每个修改动作绑定最新 stateId，坐标与 windowRef 只对最新观察有效。
• 原始 PNG 保持不变；超过附件上限时进行像素无损重压或无损切片，并记录原图/像素/图块 SHA-256。
• 独立桌面历史上限，输入文本与启动参数仅在报告中记录长度和 SHA-256。
• Windows、macOS、Linux GitHub Actions 矩阵；Windows/macOS 运行原生截图验收。

默认行为

Computer Use 默认关闭。关闭时不注册工具、不添加系统提示、不截图、不调用视觉模型，因此普通文字会话不增加插件 Token 开销。

验证

• 单元测试：64/64
• macOS 原生 observe：通过
• Windows 原生 observe：通过（GitHub Actions）
• Browser 自动化回归：通过
• npm audit：0 vulnerabilities

安装包 SHA-256：1e4d402aa15d59a6706635953323bede444d09a3fc11f50cf3a16a48358e4725

DeepSeekEyes v0.2.0

• Merge Browser Computer Use into main with Playwright-based open, observe, semantic action, wait, assertion, reporting and close operations.
• Return a fresh content-addressed screenshot, DOM-derived refs, diagnostics and stale-state-bound stateId after every browser step.
• Add native Harness settings for browser enablement, installed Edge/Chrome selection, executable path, locale, timing, viewport and observation limits.
• Raise the default base/clarification budgets to 16,384/8,192 tokens.
• Add 8K, 16K, 32K, 64K and 128K suggestions, unrestricted provider-managed output, and unbounded safe-integer custom token values.
• Omit maxTokens entirely in unrestricted mode so the selected model/Provider owns its effective output ceiling.
• Make Browser Computer Use opt-in by default so ordinary text/vision sessions receive no Browser tool or prompt overhead.
• Process only images introduced in the current request segment; historical images make zero automatic vision calls and are represented by a bounded number of compact hash-bound references.
• Compact historical Browser DOM/OCR states to a bounded recent window instead of repeatedly forwarding every full state and screenshot.
• Preserve original image events and attachment bytes while replacing processed image blocks on the model-facing session Surface, allowing direct selection of native text-only models such as DeepSeek V4 Flash.
• Add the session-scoped deepseekeyes_look tool so a native text model can request one precise reread from the preserved original attachment without enabling visual overhead in unrelated sessions.
• Fit final-model output to the resolved context window and retry one provider-reported context overflow with the exact safe capacity.
• Apply browser settings live, close stale sessions after launch-setting changes, and keep the existing final-model/vision-model GUI routing.
• Preserve screenshots and JSON reports under the configured DSH evidence directory and reject stale actions before page mutation.

DeepSeekEyes v0.1.0

首个可测试版本：在同一 DeepSeek Harness 对话中，用已配置且通过主动视觉检测的多模态模型读取原始图片，并支持 DeepSeek 向视觉模型继续追问细节。\n\n安装包 SHA-256：e56749b7d8ee37d4850ca1ebb4098a65a49c4dcbbb215b02fb6521a432ebd43a