tokenless v0.7.1

0.7.1

Patch Changes

• ce35e5a: Add the Jev smart-router engine (powered by the TypeSafe SDK) as an optional task router. When router.enabled is set with router.engine: "jev", router.providers, and a router.jevApiKey, Tokenless can route a prompt to the best-suited configured provider based on each provider's declared suitableTasks. Adds a dashboard Routing view and /dashboard-api/v1/jev/* endpoints to run routing decisions and inspect their history.
• ce35e5a: Restore the missing Lovable and Agnes provider icons (assets/providers/lovable.svg, assets/providers/agnes.svg) so both providers render correctly in the dashboard capability matrix and README provider tables.

tokenless v0.7.0

0.7.0

Minor Changes

• df65420: Replace the Dashboard overview with persisted provider, capability, outcome, execution-mode, and output-savings analytics.

• ab73799: Allow API proxy execution mode preference to hold both browser and direct, tried in the configured order when a request does not pin one. Existing config files with the older single-mode executionMode value are read as that one mode, preserving the previous enabled setting.

• a2dbf2f: Ship the initial SQLite schema migration with Tokenless API and Tokenless Harness. Upgrade the current unversioned database without replacing existing tables or data, record schema versions, and fail atomically on unsupported schemas.

  随 Tokenless API 和 Tokenless Harness 发布初始 SQLite schema migration。在不替换已有表或数据的情况下接管当前无版本数据库，记录 schema 版本，并在遇到不支持的结构时原子失败。

• 707a2e2: Add the local Spark MLX Front Door and semantic router engine.

• 228e8f2: Add an experimental Monica browser provider. tokenless/monica submits prompts on monica.im through the visible composer (Enter-key submission) and reads assistant answers from the rendered markdown surface.

• 8b215fd: Move conversation ordering to Tokenless Harness, remove Tokenless API chat/profile exclusion and automatic submitted-timeout replay, and retire unused CLI configuration and compatibility surfaces. Conflicting browser configuration and headless blockers now fail without automatically restarting the resident browser.

  将 conversation 顺序控制交给 Tokenless Harness，移除 Tokenless API 的 chat/profile 独占和提交后超时自动重发，并删除无效的 CLI 配置及兼容入口。浏览器配置冲突与 headless blocker 现在会明确失败，不再自动重启 resident browser。

• e014747: Unify Tokenless API updates across the CLI and macOS menu. Check the matching installation channel, preserve user configuration, and verify the newly installed daemon and database migration. Require confirmation for updates and support explicitly selected local release packages.

  统一 Tokenless API CLI 与 macOS 菜单的更新入口。检查对应安装渠道，保留用户配置，并验证新安装的 daemon 和数据库 migration。更新需确认，并支持显式选择本地发布安装包。

Patch Changes

• 0b40beb: Honor the selected managed profile in API requests so Terminal-Bench parent and child runs use the same provider router.

  Allow the existing bounded tool-response correction to recognize protocol headers formatted with JSON whitespace.

• 9b1aacc: Record observed provider model and effort controls at submission time and expose bounded action evidence for benchmark reports.

• ca97a53: Bundle matching Tokenless API skills with npm and macOS releases, synchronize complete skill directories during install/setup/upgrade, and add a skill-only sync command with full doctor verification. Update the installation skill and shell installer to the current setup workflow.

  随 npm 与 macOS 发布包分发配套 Tokenless API skills，在 install/setup/upgrade 中同步完整目录，并增加独立同步命令与 doctor 完整性检查。更新安装 skill 和 Shell 安装器以使用当前 setup 流程。

• 2e6116e: Use numeric date ranges and compact provider and capability icons with accessible name tooltips in Dashboard analytics.

• f099c29: Read ChatGPT thinking levels from the current account's power slider so Plus and Pro accounts expose and select their own available levels.

• bae4d71: Select ChatGPT Latest and thinking effort through the current composer menu, and prevent prompt submission in Work conversations.

• d05a81c: Support the ChatGPT Free chat composer when the newer Chat menu is absent.

• 3b4ea36: Dismiss ChatGPT's conversation-history access warning without treating it as a model generation limit, allowing new Chat prompts and responses to complete.

• 616f00d: Update ChatGPT model and thinking controls, and record the model identifier exposed on each assistant response.

• d3d7321: Show every provider and individual capability in the Dashboard usage matrix, with compact cells, date-aware titles, and simpler English and Chinese help.

• dff6809: Use persisted internal rate-limit rules for browser admission, Router fallback, and the Dashboard, with rolling-window usage and editable provider/request-type limits.

• 9d3660c: Dispatch GitHub Copilot agent CLI tasks through its native cloud Agent when GitHub repository context is supplied. Require repository selection before submission and reject local Harness execution for explicitly selected GitHub Copilot agent tasks. The delegation entry point remains experimental pending fresh real-provider acceptance.

• 8e43fc4: Show the main Dashboard charts first, remove Provider Panorama and text insights, and move chart explanations into accessible help tooltips.

• 5f8e099: Align Dashboard header controls in one row, unify icons and values, and show token and status explanations in accessible tooltips.

• d3b000f: Remove Puter from the supported provider catalog.

• 8bcb195: Remove GLHF, TheB.AI, Fenay AI, and enterprise-only Blackbox AI from the supported provider catalog.

• 6b33f7b: Fix reused Gemini conversations so each response read returns the current turn instead of a previous reply.

  修复复用 Gemini conversation 时的 response 读取，确保每次返回当前 turn，而不是之前的回复。

• af2665c: Repair Windows source-linked Tokenless API setup by invoking npm without a command shim, resolving pyenv-win's real uv executable, hiding persistent background process windows, draining CLI failures cleanly, and diagnosing native profiles through their configured browser.

  修复 Windows 源码链接安装：不再通过命令 shim 调用 npm，解析 pyenv-win 的真实 uv 可执行文件，隐藏持久后台 daemon 和 service 子进程窗口，让 CLI 失败路径正常排空句柄，并通过已配置的浏览器诊断 native profile。

• 1b5664e: Expose GitHub Copilot Ask/Agent modes, repository selection as native Projects, live model access requirements, and message token/account AI credit usage through the CLI and API.

  通过 CLI 和 API 提供 GitHub Copilot Ask/Agent 模式、映射为原生 Project 的 repository 选择、实时模型权限、回复 token 及账户 AI credits 用量。

• 5a75264: Add an experimental GitHub Copilot browser adapter with model selection, file uploads, conversation continuation, verified Harness Markdown tool round-trips, and draft attachment cleanup.

  新增实验性 GitHub Copilot 网页适配器，支持模型选择、文件上传、连续对话、已验证的 Harness Markdown 工具往返和附件草稿清理。

  Clarify that inline prompt files and files uploaded through the chat composer are separate inputs.

  明确区分提示词中的文件内容与通过聊天输入框上传的附件。

• 100034e: Reclaim idle browser work tabs with configurable retention and per-profile capacity, protect unfinished work, restore saved conversation URLs on continuation, and expose bilingual System settings and counters.

• 394072d: Remove the per-profile work-tab cap so active browser tasks can run concurrently; continue reclaiming idle tabs after the configured retention window.

• 9d3660c: Add experimental Lovable Browser mode for project creation and same-project follow-up through the selected managed profile. Show configured profile labels as display names in the Dashboard.

• 17b9f7b: Bind managed Chromium profiles to configured browser colors through Chromium's native autogenerated-theme support.

  通过 Chromium 原生自动主题支持，为 managed Chromium profile 绑定配置的浏览器颜色。

• 74fd2bd: Add an English and Chinese Dashboard rate-limit table with provider, request-type, and rule-state filters, source dates, explicit unknown coverage, and proposed ChatGPT limits clearly marked as not enforced.

• 001fce5: Reduce Dashboard rate-limit history reads so large conversation records do not repeatedly block the local API.

• d065484: Keep Windows browser version and daemon build-info checks hidden so refreshing the Tokenless API tray or Dashboard does not flash terminal windows.

  隐藏 Windows 浏览器版本检查和 daemon 构建信息检查的控制台窗口，避免刷新 Tokenless API 托盘或 Dashboard 时终端反复闪现。

  Improve the source-linked Windows tray with live task and browser-profile status, clear native notifications, resilient startup refresh, and graceful single-instance replacement.

  改进连接本地源码的 Windows 托盘：实时显示任务与浏览器 Profile 状态，提供清晰的原生通知，并增强启动刷新和单实例平滑替换的可靠性。

• 3b6eae6: Reclaim unused Tokenless API work tabs after the idle timeout, including provider homepages and tabs left by failed, canceled, or input-only jobs. Active tasks and user pages remain protected, user interaction resets the idle timer, and one blank tab keeps the browser running.

• caaebc6: Supervise all resident registered browser profiles, restore work-tab ownership across daemon restarts, recover known persisted conversations, recheck retained pages for idle state, and expose actual and untracked page counts.

• 8a789df: Honor the OpenAI Responses store flag for the local response ledger. Default storage remains enabled; store: false disables ledger continuation for that response while ordinary job history and provider-side retention remain separate.

• c1c7600: Share the Dashboard page components, styles, and profile picker with Design Atlas so Storybook previews use the same UI as the shipped application.

• Clarify source checkout refresh commands, development-link verification, optional macOS app updates, and preservation of existing configuration and data in the bundled installation skill.

  在内置安装 skill 中明确源码刷新命令、开发链接验证、可选 macOS app 更新，以及现有配置与数据的保留要求。

• 4553135: Add experimental Agnes AI Browser Chat with visible citations, account inspection, document/image input, durable conversation reuse, and the Harness read-only attachment roundtrip. Preserve unknown Web quotas; upload Markdown bytes unchanged under a supported .txt filename.

  新增实验性 Agnes AI Browser Chat 与可见引用、账号检查、文档/图片输入、持久化续聊及 Harness 只读附件闭环；保留网页额度未知状态，Markdown 内容不变并以网页支持的 .txt 文件名上传。

tokenless v0.6.0

0.6.0

Minor Changes

• 500eeb4: Present API jobs as localized chat history with readable titles, provider identities, execution modes, token estimates, and conversation transcripts.

  将 API jobs 以本地化 chat history 展示，提供可读标题、provider identity、执行模式、token estimates 与 conversation transcripts。

• 2b30ad7: Add Arena browser-mode chat, exact Direct model choice, same-conversation continuation, grounded Direct Search with visible citations, complete Battle and Side-by-Side comparison results, Direct Image generation/editing, Direct Code website generation, Agent Mode execution, and Battle Video generation with structured visible results through the selected managed browser profile.

  新增 Arena browser-mode chat、精确 Direct model 选择、同 conversation continuation、带可见 citations 的 grounded Direct Search、完整 Battle 与 Side-by-Side 对比结果、Direct Image generation/editing、Direct Code website generation、Agent Mode 执行，以及通过选定 managed browser profile 获取结构化可见结果的 Battle Video generation。

• 371039f: Add ChatGPT, Grok, Arena, and Meta AI browser-mode image asset downloads plus G4F direct image persistence: generated PNG/JPEG/WebP bytes are verified and stored under task-, conversation-, request-, and time-scoped assets, while responses expose a relative asset reference, metadata, and SHA-256 digest. Add authenticated daemon readback at /v1/asset/{taskId}/{conversationId}/{assetBatch}/{assetFile}.

  新增 ChatGPT、Grok、Arena 与 Meta AI browser mode 图片 asset 下载能力，并加入 G4F direct 图片持久化：生成得到的 PNG/JPEG/WebP bytes 会经过验证，并按 task、conversation、request 与时间保存；response 只暴露 relative asset reference、metadata 与 SHA-256 digest。新增 authenticated daemon 读取接口 /v1/asset/{taskId}/{conversationId}/{assetBatch}/{assetFile}。

• 500eeb4: Accept one validated reference image on browser /v1/images/generations requests and route it through evidence-backed image editing.

  允许 browser /v1/images/generations requests 接收一个经过验证的 reference image，并通过有证据支持的 image editing route 处理。

• f0aa326: Add caller-controlled Page Refs so independent provider work owns independent tabs while multi-turn work can reuse one stable tab.
  Remove the E2E-only provider-task isolation branch and recover cleanly when users close managed tabs.

  新增由调用方控制的 Page Ref，让独立的 provider 工作使用独立 tab，同时让多轮工作稳定复用同一个 tab。
  移除仅供 E2E 使用的 provider-task 隔离分支，并在用户关闭托管 tab 后安全恢复。

• 0ebf5e6: Expose execution mode on task capability bindings and bump the catalog schema to v3 so Browser evidence is not confused with future direct bindings.

  在 task capability bindings 中公开 execution mode，并将 catalog schema 升级到 v3，避免把 Browser evidence 与未来的 Direct bindings 混淆。

• 500eeb4: Add direct ChatGPT image generation through the canonical image endpoint with mode-aware capability routes and scoped asset persistence.

  通过 canonical image endpoint 新增 direct ChatGPT image generation，并提供按 execution mode 区分的 capability routes 与 scoped asset persistence。

• 2b30ad7: Add a setup-managed, pinned private GPT4Free provider service behind the authenticated daemon API, with native or G4F backend selection, scoped HAR/Cookie/browser auth contexts, provider/media APIs, and retained native ChatGPT and Perplexity adapters for A/B rollout.

  在 authenticated daemon API 后新增 setup-managed、pinned 的 private GPT4Free provider service，支持 native 或 G4F backend 选择、按范围隔离的 HAR/Cookie/browser auth contexts 与 provider/media APIs，并保留 native ChatGPT 与 Perplexity adapters 以支持 A/B rollout。

• f5cc843: Route direct G4F traffic through the standard OpenAI Chat Completions and Responses endpoints with upstream streaming, and remove the public G4F-specific daemon routes.

  让 direct G4F traffic 通过标准 OpenAI Chat Completions 与 Responses endpoints，并支持 upstream streaming；移除公开的 G4F-specific daemon routes。

• 77edff2: Add the durable tokenless agent run|read|resume|cancel workflow with authenticated daemon routing, local stdio MCP catalogs, exact approval arguments, and restart-safe provider continuation.

  新增 durable tokenless agent run|read|resume|cancel workflow，支持 authenticated daemon routing、local stdio MCP catalogs、精确 approval arguments，以及可安全跨 restart 继续的 provider continuation。

• 610136f: Replace the ad hoc coding-prompt collection with a pinned FeatureBench agent runtime: task-scoped provider channels, container filesystem/shell/test tools, official fb eval verdicts, non-overwriting run reports, and real-provider wiring/full E2E entry points.

  用 pinned FeatureBench agent runtime 替代临时 coding-prompt collection：提供 task-scoped provider channels、container filesystem/shell/test tools、官方 fb eval verdicts、不会覆盖旧结果的 run reports，以及真实 provider wiring 与完整 E2E entry points。

• bb66332: Expose the consolidated G4F vendor catalog as Direct providers and show each provider's supported execution modes in the dashboard.

  将 consolidated G4F vendor catalog 公开为 Direct providers，并在 dashboard 中展示每个 provider 支持的 execution modes。

• dc9f8fd: Add a bilingual Experimental Router area to Providers with persisted provider routing roles, profile-aware eligibility, Chrome Prompt API testing and compatibility diagnostics, plus focused provider detail pages for routing and existing controls.

  在 Providers 中新增双语 Experimental Router 区域，支持持久化 provider routing roles、profile-aware eligibility、Chrome Prompt API testing 与 compatibility diagnostics，并提供聚焦 routing 和现有 controls 的 provider detail pages。

• abf8626: Add experimental Meta AI browser chat, file upload, and Instant/Thinking controls through the selected signed-in managed profile.

  通过选定的 signed-in managed profile 新增 experimental Meta AI browser chat、file upload 与 Instant/Thinking controls。

• 81709ce: Accept modern OpenAI function tools and structured final output on Chat Completions requests, including terminal SSE frames, complete multiple-call history, every tool_choice mode, parallel_tool_calls, strict: true arguments, json_object, and a published closed-object json_schema subset. Tokenless validates request-scoped catalogs, call/result history, tool arguments, and final JSON before returning model-ordered standard calls or exact schema-valid JSON text. Structured JSON numbers use one canonical finite spelling and safe integers so returned text cannot differ from the AJV value. Structured output also works without tools and after caller-owned tool execution; the API proxy never executes tools. Invalid requests fail before provider submission, while invalid provider framing, calls, arguments, or structured content fail with provider_output_protocol_error. Prompt-emulated calls use one strict whole-response JSON object, accepting only bare JSON or one complete json/text fence; marker wrappers, prose, and multiple fences fail closed. A nonce-correlated outer kind: final JSON string-escaping failure may receive one bounded same-provider correction before exposure; corrected inner structured content must still pass strict parsing and schema validation.

  The Universal API and Standalone Web Agent Harness now share only strict JSON parsing and AJV 2020 setup. Their OpenAI choice/correction/structured-output and Harness action_batch/execution semantics remain separate.

  Universal API 与 Standalone Web Agent Harness 现在只共享 strict JSON parsing 与 AJV 2020 setup；两者的 OpenAI choice/correction/structured-output 以及 Harness action_batch/execution semantics 仍然独立。

  Add the Universal Responses API on /v1/responses and /v1/openai/responses. It supports current flat function tools, typed call/result items, structured text.format, non-streaming output, typed terminal streaming, full-input replay, and same-route previous_response_id continuation through a 24-hour, 1,000-entry local ledger. Missing, expired, mismatched, or unverifiable opaque replay state fails before provider submission; Tokenless stores no fabricated reasoning or provider session state.

  在 /v1/responses 与 /v1/openai/responses 新增 Universal Responses API，支持当前 flat function tools、typed call/result items、structured text.format、non-streaming output、typed terminal streaming、full-input replay，以及通过 24 小时、1,000 条目的 local ledger 实现同 route 的 previous_response_id continuation。缺失、过期、不匹配或无法验证的 opaque replay state 会在 provider submission 前失败；Tokenless 不存储伪造 reasoning 或 provider session state。

  Add explicit tokenless/auto for OpenAI browser-mode tool and structured-final requests. Auto filters enabled and currently usable providers by the complete evidence-backed requirement set, preserves canonical tool history and versioned public call ids across providers, prefers the prior auto call's settled provider when still eligible, and reuses the existing durable pre-submission fallback plan. Responses ledger continuation treats its settled provider as portable affinity. The response metadata reports the actual settled provider, prompt strategy, and redacted provider attempts; exact provider models never auto-switch, and post-submission outcomes never replay to another provider.

  为 OpenAI browser-mode tool 与 structured-final requests 新增显式 tokenless/auto。Auto 会按完整的 evidence-backed requirement set 筛选已启用且当前可用的 providers，跨 provider 保留 canonical tool history 与 versioned public call ids；此前 auto call 的 settled provider 仍符合条件时会优先复用，并沿用现有 durable pre-submission fallback plan。Responses ledger continuation 将 settled provider 视为 portable affinity；response metadata 报告实际 settled provider、prompt strategy 与 redacted provider attempts，exact provider models 不会自动切换，post-submission outcomes 也不会重放到其他 provider。

• e3ce67e: Show Browser and Direct capability badges for every provider and add profile-scoped execution-mode controls to provider details.

  为每个 provider 展示 Browser 与 Direct capability badges，并在 provider details 中新增 profile-scoped execution-mode controls。

• b5ac343: Add one authenticated /v1/images/generations endpoint for browser and direct image generation, route CLI image runs through it, remove the public G4F-namespaced image route, and close Gemini, Dola, and Doubao browser image routes.

  为 browser 与 direct image generation 新增统一的 authentica...

tokenless v0.5.0

0.5.0

Minor Changes

• f89070a: Add an opt-in Codex integration with reversible global guidance, native lifecycle hooks, exact project/chat/turn/tool-call binding, App Server enrichment, and durable provider conversation continuity.

• 1347720: Enable output savings measurement by default, durably process first-response tokenization after provider completion in an isolated child process, and keep a prominent savings summary on the dashboard with a clear disabled state.

• e840aee: Add Dola as an experimental managed-browser provider with signed-in chat, Fast/Pro selection, file input, image/writing/video/translation/homework capability candidates, navigation, sanitized-DOM capture, and real-provider release gates.

  新增实验性 Dola managed-browser provider，覆盖登录态 chat、Fast/Pro、file input、图片/写作/视频/翻译/作业能力候选、navigation、sanitized DOM capture 与真实 provider release gates。

• 3eb3f73: Add native Google Chrome and Brave selection after the Anti-Detect setup choice, support a verified user-supplied executable path, attach through the selected browser's managed CDP endpoint, and retire browser-profile import compatibility checks. Missing native browsers now produce an end-of-setup action warning; the first browser action validates the configured path or standard installation before use.

• 6075a7f: Replace managed browser profile imports with headed native connections to the user's running Google Chrome 144+ instance. After the user enables Remote Debugging, Chrome manages the CDP endpoint and Tokenless discovers it without a fixed-port setting. Tokenless checks support by connecting, manages only its own tabs, and disconnects without closing Chrome when the daemon stops.

• e0c7f3c: Store complete per-profile provider configuration in config.json under profiles, migrate the legacy side table with the registered profiles, and expose profile configuration directly in CLI and dashboard payloads.

• 15ebb8c: Separate provider-native workflows from canonical capabilities, remove skill.invoke, and deliver caller-selected Harness Skills through the standard chat and file-upload route.

  将 provider-native workflow 与 canonical capability 分离，删除 skill.invoke，并通过标准 chat 与 file-upload route 交付 caller-selected Harness Skills。

• 1c0d2e9: Add explicit tokenless setup --install-codex onboarding and bind Codex CLI calls to concrete CODEX_THREAD_ID tasks while preserving Hook session-tree provenance and automatically reusing native provider workspaces.

Patch Changes

• 75ad601: Consolidate managed browser control on CDP, remove the obsolete browser connection mode from configuration and UI contracts, and keep Playwright's page and locator APIs for automation. Existing configuration files must remove browserConnectionMode; all legacy values are rejected.

• 1bf51f9: Add checksum-pinned official CloakBrowser downloads for Intel macOS and Linux arm64/x64 while retaining the existing Apple Silicon macOS and Windows x64 pins.

• f1ea0fb: Advertise experimental DeepSeek chat and Markdown file-upload routes after a real headed managed-browser run proved three visible attachment cards, prompt submission, and response reading.

  在真实 headed managed browser 运行验证三张可见附件卡片、prompt 提交与回复读取后，公开实验性的 DeepSeek chat 和 Markdown 文件上传 routes。

• cccf099: Open the loopback local console directly without a bootstrap ticket while preserving UI CSRF and machine API bearer protections.

• dbb999c: Report configuration completeness in tokenless doctor, including stale or missing user-supplied Chrome and Brave executable paths.

• f1ea0fb: Update Dola's relative upload and submit controls and recognize physical Markdown attachment cards, while leaving chat and file-upload routes unadvertised because the selected profile accepted only one visible document card.

  更新 Dola 的相对 upload 与 submit controls，并识别物理可见的 Markdown 附件卡片；由于选定 profile 仅接收一张可见文档卡片，chat 与 file-upload routes 仍不公开。

• 562baed: Recognize Gemini's signed-in navigation and Z.ai's visible user menu before reporting a guest session from auth.status.

• a26661a: Add experimental Gemini Markdown file upload with physical attachment-card acceptance and the visible Upload & tools flow. Implement Qwen's visible Upload attachment chooser and Markdown card detector while keeping its route unadvertised because the real built-product send click had no visible transition and submission remained ambiguous.

  新增实验性的 Gemini Markdown 文件上传，通过物理附件卡片与可见 Upload & tools 流程证明接收。实现 Qwen 的可见 Upload attachment chooser 与 Markdown 卡片检测；由于真实 built-product 的发送点击没有可见转换，submission 仍存在歧义，继续不公开其 route。

• 94bf3c1: Keep hook-injected project and chat names as task identity unless Workspace handling is explicitly requested.

• 7871659: Keep managed browsers resident across jobs and daemon restarts. Dashboard provider-readiness refreshes run headlessly when the profile is idle, reuse an existing headed browser without replacing or foregrounding it, and reserve visible browser handoff for explicit user actions.

  Managed browsers 会跨 jobs 和 daemon 重启保持常驻。Dashboard 刷新 Provider 就绪状态时，会在 Profile 空闲时以 headless 方式运行；已有 headed browser 时则静默复用且不替换或带到前台，仅在用户显式操作时启动可见 handoff。

• ed476a9: Upload the Harness System Prompt and every selected Skill as independent Markdown files in one visible provider batch.

• f1ea0fb: Reconnect a managed browser that was already running before resident-browser session metadata existed, instead of trying to launch a second browser against the same profile.

  重新连接在驻留浏览器 session 元数据出现前已运行的托管浏览器，避免对同一 profile 启动第二个浏览器。

• c3a7c2d: Refresh Tokenless skills in the canonical shared directory and existing direct agent skill directories during setup and upgrade, including Codex and Claude Code.

• 913b9d0: Let production managed browsers use native credential storage so compatible copied profiles can preserve macOS Keychain-encrypted sign-in state, and document the expected Safe Storage prompt.

• f1ea0fb: Perplexity file upload remains unadvertised when the selected Free plan visibly requires an upgrade for a third document.

  当选定的 Free plan 对第三个文档显示可见升级要求时，Perplexity 文件上传继续不对外公开。

• 0290c36: Move browser opening from the global dashboard header to profile-specific controls, always open those controls in a visible managed browser, and let headed dashboard pages respond to the managed browser window size.

• 83b6352: Use profile slugs as the only profile names across the CLI and local console.

• f1ea0fb: Recognize Claude attachment filenames before concatenated card metadata and Kimi filenames with separate extension badges, while retaining count-plus-extension evidence. During the CDP-only transition, accept only the exact legacy browserConnectionMode: 'playwright' value and omit that retired field from normalized config.

  正确识别 Claude 附件卡片中拼接元数据前的文件名，以及 Kimi 将扩展名拆分为独立标记的成功文件卡片，同时保留数量加扩展名证据。迁移到 CDP-only 控制期间，仅接受精确的旧值 browserConnectionMode: 'playwright'，并从规范化配置中省略该已退役字段。

• 6148137: Keep headed provider tabs and the local dashboard open when another provider tab is closed or the dashboard is refreshed.

• 0290c36: Simplify interactive setup provider selection with an all-enabled list and numbered removals.

• 5db3fdd: Wait for Qwen attachments to finish visible parsing before prompt submission, use provider-specific interaction windows, require provider acknowledgement after one actionable submit click, and fail response reads that have no visible answer.

  在提交 prompt 前等待 Qwen 附件完成可见解析，使用 provider-specific 交互等待窗口，在唯一一次可操作提交点击后要求 provider acknowledgement，并让没有可见答案的 response read 明确失败。

• dfd367c: Add an Overview refresh control that rechecks sign-in readiness for every enabled provider in the selected managed profile.

• f129343: Stop rejecting managed profile registries based on POSIX mode bits so valid Windows profiles remain usable while Tokenless continues to create registry files with private permissions where supported.

  不再根据 POSIX mode 位拒绝托管浏览器配置注册表，避免有效的 Windows 配置被误判；Tokenless 仍会在平台支持时以私有权限创建注册表文件。

• 091258f: Persist bounded, text-free response selector diagnostics with sanitized snapshots and response reads.

• 83b6352: Stop the daemon only after background output-savings work has released the shared database, and let live release gates inspect an already-resident configured browser.

  等待后台 output-savings 工作释放共享数据库后再停止 daemon，并允许实时发布门禁检查已常驻的配置浏览器。

• 0480d5b: Dismiss blocking Z.ai announcements before prompt entry, reuse open Doubao mode menus, wait for hydrated provider choice controls, preserve mapped native Project conversations, read complete Claude responses, apply Claude Project instructions through the visible instructions dialog, and stop advertising incomplete DeepSeek and Qwen conversation, Grok and Kimi native Project, Qwen mode, and Kimi model, library, or long-running workflows.

  在输入提示词前关闭阻塞操作的 Z.ai 公告、复用已打开的豆包模式菜单、等待 provider choice 控件完成加载、保留已映射的原生 Project 对话、完整读取 Claude 回复并通过可见的说明对话框应用 Claude Project instructions，同时停止宣称未闭环的 DeepSeek、Qwen 对话、Grok、Kimi 原生 Project、Qwen mode 以及 Kimi model、library 和 long-running workflows。

• e177aef: Require a stable visible provider response before a completed run is returned.

• 35343ce: Check provider readiness in batches of up to three task-owned background tabs and close every temporary tab after observation.

  Provider 就绪检查会以最多三个一批的方式使用任务自有后台 tab，并在观察结束后关闭所有临时 tab。

• 47b7373: Wait for provider responses and managed job completion without a universal elapsed-time deadline while preserving explicit cancellation and runner shutdown. Explicit --timeout-ms continues to cancel the waiting job when it elapses.

  等待 provider 响应和托管任务完成时不再使用统一的耗时截止，同时保留显式取消与 runner 停止能力。显式 --timeout-ms 到期时仍会取消正在等待的任务。

• e0c7f3c: Use stable locale catalog keys for CLI output and local-console error summaries, with Simplified Chinese privacy guidance.

• 15ebb8c: Accept multi-file provider uploads from newly visible attachment evidence by count and file-extension multiset, even when providers rename files.

  通过新增可见附件证据的数量与文件扩展名多重集接受多文件 Provider 上传，即使 Provider 重命名文件也...

tokenless v0.4.1

0.4.1

Patch Changes

• b4556a9: Fix tokenless upgrade --json rejecting its default empty capability argument state before starting the upgrade workflow.

tokenless v0.4.0

0.4.0

Minor Changes

• f676cd6: Add experimental signed-in Kimi chat, text-file upload, Web search with citations, model and thinking-effort controls, and provider-native Plugin and Skill controls with real Cloak capability gates.
• 2986940: Add profile-bound browser runtime selection with system-first auto discovery, a locked managed Chrome for Testing fallback, explicit Anti-Detect CloakBrowser setup, safe known-Chromium profile/version inventory before Cloak download, exact four-component Cloak compatibility classification, interactive or explicit non-interactive clean-profile confirmation, clean runtime-isolated profiles, exact daemon runtime errors, transactional artifact security verification, per-provider setup review tabs, keychain-neutral E2E Cloak launches, separate doctor profile/runtime compatibility checks, and gated real-browser acceptance coverage.
• bf24d96: Cache the verified browser executable path in persistent config, resolve concrete browser selections during setup and install, retry standard-path discovery after stale cache entries, and add CLI and dashboard recovery controls for custom system-browser paths. Interactive setup no longer asks users to choose a normal browser runtime. Anti-Detect setup discloses Cloak installation in its initial question and uses one clean-or-compatible-profile source choice, without redundant import, copy-consent, or final installation prompts. Opaque import remains limited to profiles whose Chromium version exactly matches the supported CloakBrowser runtime, with a final compatibility check at the copy boundary.
• f6dd951: Add experimental four-provider capability inspection, explicit Workspace fallback modes, trusted conversation continuation, and visible file-upload acceptance evidence.
• 1d44047: Add a canonical task-capability catalog, expose evidence-backed provider routes through tokenless capabilities list, and route normal runs by explicit and inferred capability requirements before browser mutation.
• 3f21056: Rank automatic provider fallback by the run's complete capability requirements, recheck live eligibility before mutation, reject under-declared routes, preserve a versioned context envelope, expose structured durable fallback evidence, and fail closed when file selection lacks visible provider acceptance.
• ef33467: Unify and simplify the local CLI-daemon HTTP boundary under one Tokenless Daemon API v1 OpenAPI contract: keep only public job and browser-runtime routes, run Playwright coordination in-process, require proof before bearer use, and remove unreleased legacy payloads and extension compatibility.
• d099d88: Add experimental Qwen guest-session support through the shared provider registry and visible Playwright workflow.
• 09f5f07: Add the experimental DeepSeek visible-session provider with safe navigation, sign-in detection, baseline chat controls, and evidence-gated capability classification.
• 3f21056: Add experimental DeepSeek visible provider support with exact Instant, Expert, and Vision mode selection, independent DeepThink and Search controls, hCaptcha detection, mode-aware capability inspection, and provenance-bound real-session DOM fixtures.
• b9421d3: Add an experimental signed-in Doubao adapter with visible account and free-tier inspection, Cloak-tested file upload, exact mode and coding-oriented Web skill controls, visible unavailable-state handling, prompt and response handling, and provider challenge detection.
• 50a20b5: Add the secure bilingual local web control plane, verified browser discovery and executable-path recovery, explicit opaque local profile import and re-import, profile-scoped provider and browser preferences, reserved dashboard tabs, setup handoff, dashboard CLI command, and browser-facing administration API.
• d3d78ce: Add English and Simplified Chinese CLI output selected during setup or through the persistent language config.
• e3c1ca0: Add opt-in output savings measurement with a lazily downloaded checksum-pinned local tokenizer, durable per-job attribution and aggregates, explicit CLI and bilingual dashboard controls, doctor diagnostics, and a real cross-platform runtime verification workflow.
• 3f21056: Add experimental Perplexity guest chat routing with live prompt submission, completed responses, normalized citations, and visible citation-link closure through a managed Cloak profile.
• bd4b7c9: Route implicit provider jobs through cached setup observations, report pre-submission errors when no cached provider is usable, and make doctor report provider usability plus normally stopped on-demand runtime state without treating either as damage.
• 3f21056: Add a packaged subscription-aware provider rate-limit catalog, durable submission history, cadence-based admission and deferral, in-scope provider fallback, visible cooldown handling, and tokenless limits inspect diagnostics.
• 2986940: Automatically fallback implicit runs across capability-compatible providers before human handoff while preserving one durable job and provider-attempt history.
• b9421d3: Rename persistent preferred provider configuration to providerWhitelist, default it to every non-disabled provider except Gemini, and migrate existing config on the next write.
• bb7bf16: Ship Tokenless as a pure JavaScript package with the TypeScript daemon embedded in the CLI package, removing platform-specific native runtime packages and standalone Playwright runner artifacts.
• 09f5f07: Preserve independent browser tabs for different providers, projects, and conversations unless a local job explicitly requests replacement.
• 7df255f: Add concise human-readable CLI output by default, --verbose diagnostics, and cross-platform --color/--no-color controls without changing JSON output.
• 7652158: Add Qwen-specific mode discovery and selection, including Deep Research variants and Auto/Thinking/Fast effort controls.
• 2f0f90f: Add real Claude and Grok native Project creation, exact reuse, instruction outcomes, and durable Project-scoped conversation identity. Use the canonical Qwen Studio chat surface with hydration-safe guest submission.
• df5edf7: Fail provider authentication checks closed on visible account controls and include visible provider usernames and subscription labels in managed profile output.
• 3f21056: Add experimental Z.ai guest chat routing with Cloak-verified prompt submission, completed visible responses, conversation mapping, and durable state.

Patch Changes

• f676cd6: Keep headed automation tabs and navigation in the background, foregrounding the browser only for explicit user handoffs.
• f6dd951: Add case-sensitive -P profile and -p provider CLI shortcuts.
• 107ee7e: Add an experimental config-only Playwright/CDP browser connection mode and a two-mode real-browser capability matrix while preserving the existing default.
• 28cfb6d: Reorganize CLI help into concise canonical and advanced command groups.
• bacffef: Classify Claude Cloudflare human checks as a narrow live E2E release-gate known issue when the durable blocker code is structured as visible_cloudflare_turnstile or visible_cloudflare_interstitial.
• 21844cc: Publish a complete bilingual CLI command reference and link it from built-in help and public READMEs.
• 350fba0: Start the local daemon on demand with SQLite-backed dynamic endpoint discovery, recover leased and checkpointed jobs after restart, and add agent-scoped one-time outcome replay with durable full-result lookup.
• 0d438ec: Accept non-empty visible composer content after prompt input instead of requiring an exact DOM text match.
• f6dd951: Report Grok profiles as Free or SuperGrok from visible model entitlements and inventory redacted provider DOM fixtures by page URL.
• 0dc3e49: Add provider access/tier reporting and a provider-session state machine that runs ChatGPT and Gemini guest sessions, hands unauthenticated Claude and Grok jobs to the user before task input, and persists redacted structural DOM snapshots.
• 2986940: Preserve limited Chromium compatibility state when importing a browser profile with explicit user consent.
• 7df255f: Include Gemini in new default provider profiles and write real-provider E2E results as provider-grouped capability reports.
• b9421d3: Keep every production Chromium launch keychain-neutral with --password-store=basic and --use-mock-keychain, regardless of the configured browser executable or connection mode.
• 0d438ec: Keep managed-profile initialization user-run instead of triggering it from agent skills.
• f676cd6: Reuse an already approved provider page instead of refreshing it before every action, preserving visible conversation and control state across CLI jobs for the same task.
• 378ed4c: Keep one stable browser instance per active managed profile while removing the macOS-specific inspection launcher.
• 0d438ec: Wait for visible prompt input and submit controls, distinguish availability timeouts from action failures, and allow free or anonymous prompt surfaces without an authentication precheck.
• 1d44047: Allow tokenless profiles open to launch a managed profile without selecting or navigating to a provider unless --provider is passed.
• abc58bb: Record intermittent Qwen DNS failures as retryable suspected rate limits without claiming confirmed provider throttling, and document the distinction and recovery steps.
• f676cd6: Wait for Qwen's visible Deep Research variant menu to settle and scroll the selected option into view before clicking it.
• bcabe9a: Gracefully restart a proof-verified stale local daemon from the current CLI package when its package version or control API revision no longer matches.
• 0dc3e49: Make ordinary daemon reuse depend on verified ready identity plus the exact package version, and let tokenless setup reconcile same-home version drift only after proof verification.
• df5edf7: Check each provider's setup sign-in status once and report it without opening a login handoff or retrying.
• 0d43...

tokenless v0.3.0

0.3.0

Minor Changes

• 462123b: Add browser visibility policy support to Tokenless config and managed Playwright job requests, including durable resume after a headless job parks for user handoff.

Patch Changes

• 6d52df5: Ship the current Tokenless reliability release as one patch:

  • Archive the browser extension and Native Messaging host under legacy/, leaving the managed Playwright daemon as the active runtime.
  • Remove the experimental direct runtime, direct broker, account/project routing commands, direct public exports, and direct-only documentation so managed Playwright through the authenticated local daemon is the only execution path.
  • Provision and verify the local Rust daemon before provider sign-in, restart only process-correlated stale daemons, and keep doctor diagnostics read-only.
  • Add prompt-free tokenless upgrade with concise human progress and structured --json output to update the global CLI, refresh Tokenless agent skills, reconcile the packaged daemon through the verified new CLI, and report its final doctor result.
  • Keep animated setup progress stable on one terminal line, including in narrow and hosted terminals.
  • Make setup check every supported visible provider, preserve ChatGPT as the default run provider, reject obsolete setup-only provider filters, and document the supported browser-profile import scope.

• 08000f1: Enable Chromium sandboxing for managed visible browser sessions so Chrome no longer launches with the unsupported --no-sandbox flag.