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 authenticated /v1/images/generations endpoint，让 CLI image runs 通过该 endpoint 路由；移除公开的 G4F-namespaced image route，并关闭 Gemini、Dola 与 Doubao browser image routes。

• 7653c94: Add an experimental Tokenless Harness Chrome side-panel extension with scoped Dashboard pairing, semantic page observation, full private local evidence, and individually approved input, click, submit, radio, upload, and navigation actions for one explicitly selected tab.

  新增 experimental Tokenless Harness Chrome side panel 扩展，通过 scoped Dashboard pairing、semantic page observation、完整私有本地证据，以及逐项批准的 input、click、submit、radio、upload 与 navigation 操作一个明确选择的 tab。

• 8885628: Add synchronous Tokenless Harness delegation with bounded workspace read/search tools, a reversible DeepSeek Harness subagent provider integration, and the pinned Terminal-Bench 2.0 runner for that combined lane. Document Codex as explicit delegation because its hooks do not replace native subagent execution.

  新增同步 Tokenless Harness delegation、有界 workspace read/search tools、可逆的 DeepSeek Harness subagent provider 集成，以及用于该组合 lane 的固定 Terminal-Bench 2.0 runner。Codex 因 hooks 无法替换 native subagent execution，按显式 delegation 如实记录。

• c5329ae: Add the authenticated menu bar snapshot contract, dashboard job deep links, and a read-only tokenless upgrade --check command for native macOS clients.

  新增 authenticated menu bar snapshot contract、dashboard job deep links，以及供原生 macOS clients 使用的只读 tokenless upgrade --check command。

• 5c60425: Move every non-compatibility bearer-authenticated Tokenless machine endpoint under /v1/private/*. Make packages/contracts a documentation-only OpenAPI source with one generated reference, and move runtime primitives and the provider-turn Client Adapter to their owning packages without changing payload, lifecycle, Dashboard, CLI, Harness, or provider behavior.

  将所有非 compatibility 的 bearer-authenticated Tokenless machine endpoint 统一迁移到 /v1/private/*。packages/contracts 收纯为 documentation-only OpenAPI source 与单一 generated reference，runtime primitive 和 provider-turn Client Adapter 回到各自 owner package；payload、lifecycle、Dashboard、CLI、Harness 与 provider behavior 保持不变。

• 28362bf: Let DeepSeek Harness benchmark agents inspect one file, delegate one bounded read-only workspace pass, then batch or locally search the permitted changes with a hard verification budget. Record safe live-page limit evidence, configured provider plan observations, and per-interaction token estimates in Terminal-Bench reports. Add the one-pass 89-task Terminal-Bench sweep as the required phase gate before the formal k=5 run.

  让 DeepSeek Harness benchmark agents 检查一个文件、执行一次有界的只读 workspace pass，再对获准变更进行 batch 或本地搜索，并设置严格的 verification budget。在 Terminal-Bench reports 中记录安全的 live-page limit evidence、配置的 provider plan observations 与每次 interaction 的 token estimates；在正式 k=5 run 前增加一次性 89-task Terminal-Bench sweep 作为必需的 phase gate。

  Terminal-Bench reports now distinguish successful and failed DeepSeek Harness parents from unsettled parent routing requests, and the sweep gate rejects those failures even when verifier rewards pass.

  Terminal-Bench reports 现在会区分成功和失败的 DeepSeek Harness parents 与尚未定案的 parent routing requests；即使 verifier rewards pass，sweep gate 仍会拒绝这些 failures。

• 5a5f421: Turn System into a complete configuration center with structured editing, a fresh server-backed config.json view, bottom-anchored navigation, and a dedicated Chat History icon.

  将 System 扩展为完整配置中心，提供结构化编辑、由 server 实时读取的 config.json 视图、固定在底部的导航入口，以及专用的对话历史图标。

• ef2ec65: Open Chat History conversations as complete Dashboard pages, keep provider-readiness feedback visually stable, and synchronize the local UI session before protected mutations.

  将对话历史详情改为完整的 Dashboard 页面，保持 Provider 就绪状态反馈稳定，并在受保护的修改请求前同步本地 UI 会话。

• f3426da: Add a Google Chrome Prompt API Dashboard flow for generating the pinned Terminal-Bench semantic manifest.

  新增 Google Chrome Prompt API Dashboard 流程，用于生成 pinned Terminal-Bench semantic manifest。

• 62d15e2: Localize CLI setup and success output, the full Dashboard, generated HTTP API references, and public product documentation in English and Simplified Chinese without changing commands, protocol values, or runtime behavior.

  为 CLI setup 与成功输出、完整 Dashboard、生成式 HTTP API reference 和公开产品文档提供 English/简体中文双语支持，不改变 command、protocol value 或 runtime behavior。

• bf11fbe: Keep Tokenless Harness runs in daemon memory and remove durable admissions, restart recovery, and --admission-ref.

  让 Tokenless Harness runs 保存在 daemon memory 中，并移除 durable admissions、restart recovery 与 --admission-ref。

• 2056fe7: Add provider-neutral semantic preferences and provider-scoped rate-limit fallback for Tokenless API auto routing, with per-provider Terminal-Bench observations.

  Allow auto private provider-turn continuations to retain their settled provider mapping while carrying portable provider-home fallback alternatives; exact provider bindings remain pinned.

  为 Tokenless API auto routing 增加 provider-neutral semantic preference 与 provider-scoped rate-limit fallback，并在 Terminal-Bench 中按 provider 记录观测指标。

  允许 auto private provider-turn continuation 保留 settled provider mapping，同时携带 portable provider-home fallback alternative；精确 provider binding 仍保持 pinned。

• 86d7b64: Allow every eligible text conversation provider to carry Tokenless-owned prompt-emulated function calls and structured finals. Recover one unambiguous JSON object from surrounding provider prose while preserving protocol, nonce, tool, call-count, history, and schema validation. Close a still-visible Z.ai announcement through its scoped dialog control when Escape does not unblock the real send button. Normalize provider-rendered non-breaking JSON spaces, read Meta AI JSON from its visible Raw view, and keep Arena Direct from treating the submitted prompt as an assistant answer.

  允许每个 eligible text conversation provider 承载由 Tokenless 实现的 prompt-emulated function call 与 structured final。在 provider prose 中提取唯一且无歧义的 JSON object，同时保留 protocol、nonce、tool、call-count、history 与 schema validation。当 Escape 未能解除 Z.ai 公告对真实发送按钮的遮挡时，只通过该公告内部的 dialog control 关闭它。统一 provider 页面渲染出的 JSON non-breaking space，从 Meta AI 可见的 Raw 视图读取 JSON，并防止 Arena Direct 把已提交的 prompt 当成 assistant answer。

• 315aa73: Add a dedicated setup route that discovers installed Chrome, Brave, and CloakBrowser runtimes, lets users add a verified native executable path, and saves the selected runtime with the config-wired default profile.

  新增专用 setup 路由，发现已安装的 Chrome、Brave 和 CloakBrowser runtime，允许用户添加经过验证的原生 executable path，并将所选 runtime 与配置关联的默认 profile 一起保存。

• 455a289: Use one tokenless.sqlite3 for current business records: jobs, provider Project and conversation mappings, Responses API continuation entries, output-savings events, provider status observations, and Tokenless Harness context when used. Keep only in-flight Web AI execution in daemon memory; remove custom writer/startup locks, claim leases, checkpoints, replay/resume, delayed recovery, idempotency-key aliases, dispatch-certainty state, request-cancellation tombstones, legacy profile-registry import, and the background output-savings queue. Unfinished jobs fail with job_interrupted after restart; completed history and continuation state remain available.

  只使用一个 tokenless.sqlite3 存储当前业务记录：jobs、provider Project 与 conversation mappings、Responses API continuation entries、output-savings events、provider status observations，以及使用 Tokenless Harness 时的 context。只有 in-flight Web AI execution 留在 daemon memory；移除自定义 writer/startup lock、claim lease、checkpoint、replay/resume、delayed recovery、idempotency-key alias、dispatch-certainty state、request-cancellation tombstone、旧 profile-registry import 与后台 output-savings queue。重启后未完成 job 以 job_interrupted 失败；已完成历史和 continuation state 会继续保留。

Patch Changes

• fbd9d59: Resolve Terminal-Bench Git recovery conflicts with the recovered cherry-pick version.

  使用 recovered cherry-pick version 解决 Terminal-Bench Git recovery 冲突。

• e78f317: Require Terminal-Bench Git recovery to finish cherry-pick conflict resolution before completing.

  要求 Terminal-Bench Git recovery 在完成前彻底解决 cherry-pick 冲突。

• c541f6d: Require the Terminal-Bench Git recovery parent to execute recovery through its bash tool before returning a final response.

  要求 Terminal-Bench Git recovery parent 在返回最终响应前通过 bash tool 执行恢复操作。

• ffbcadd: Make the Terminal-Bench Git recovery parent perform and verify the recovered change instead of returning instructions.

  让 Terminal-Bench Git recovery parent 实际执行并验证恢复操作，而不是只返回操作说明。

• 91a7345: Keep Terminal-Bench Harness child terminal results inside the required JSON response envelope.

  确保 Terminal-Bench Harness child terminal result 保持在所需的 JSON response envelope 内。

• 893b2c9: Guide Terminal-Bench Git recovery through existing Git metadata and make benchmark Harness delegation distinguish workspace content searches from shell commands.

  引导 Terminal-Bench Git recovery 使用现有 Git metadata，并让 benchmark Harness delegation 明确区分 workspace content search 与 shell command。

• 0029040: Accept valid strict JSON tool-protocol finals whose text contains Markdown code fences, while continuing to reject multiple outer response fences.

  接受 final text 含 Markdown code fence 的有效 strict JSON tool-protocol response，同时继续拒绝多个 outer response fences。

• 5685e9c: Disable Dola image-generation and artifact routes after the 2026-09-01 real run could not reach the daemon, while retaining verified chat and document routes.

  在 2026-09-01 真实运行无法连接 daemon 后，关闭 Dola 的图像生成与 artifact 路由，同时保留已验证的聊天和文档路由。

• 902644d: Close Arena website-generation and agent-execution routes until stable real response evidence is restored, and correct the workspace response baseline check for one visible answer.

  在稳定的真实响应证据恢复前关闭 Arena website-generation 和 agent-execution 路由，并修正 workspace response baseline 检查以匹配单个可见回答。

• 4ba3477: Fix current Arena model selection and Qwen effort selection controls.

  修复 Arena 当前 model 选择和 Qwen effort 选择控件。

• 4f89cfe: Granularize browser file-input capabilities and stabilize Qwen document uploads.

  细分 browser file-input capabilities，并稳定 Qwen document uploads。

• 3553ec6: Make local API proxy failures distinguishable and serve the default OpenAI paths. Errors now carry a real HTTP status and a stable code — 404 for an unknown model, 413 for an oversized body, 499 on client disconnect, 502 for a visible-provider failure, 503 for a disabled proxy or unready profile, 504 for the completion deadline — so clients can decide whether to retry without matching message strings. POST /v1/chat/completions and GET /v1/models are accepted as aliases of the /v1/openai routes, letting an unmodified OpenAI client work with only a base-URL change, and every proxy route now requires the proxy to be enabled. tokenless api-proxy status --json reports the additional openaiDefault endpoint.

  让 local API proxy 的失败可区分，并提供默认 OpenAI paths。错误现在带有真实 HTTP status 与稳定 code：未知 model 为 404、请求体过大为 413、client disconnect 为 499、visible-provider failure 为 502、proxy disabled 或 profile 未就绪为 503、completion deadline 超时为 504，客户端无需匹配错误文本即可决定是否重试。POST /v1/chat/completions 与 GET /v1/models 作为 /v1/openai routes 的 aliases 接受，未修改的 OpenAI client 只需更换 base URL；每个 proxy route 都要求 proxy 已启用，tokenless api-proxy status --json 也会报告新增的 openaiDefault endpoint。

• 0b98b27: Keep Doubao capability inspection from opening the native file picker.

  保持 Doubao capability inspection 不打开 native file picker。

• 3e5e203: Restore Doubao image generation on the current Seedream surface and submit control.

  在当前 Seedream surface 上恢复 Doubao image generation 与 submit control。

• b5ac343: Detect Grok Free, SuperGrok Lite, and SuperGrok subscription access from the live model menu, and stop advertising Heavy or Build as selectable when the SuperGrok upgrade boundary is visible.

  从真实 model menu 检测 Grok Free、SuperGrok Lite 与 SuperGrok 订阅权限，并在 SuperGrok upgrade 边界可见时不再将 Heavy 或 Build 报告为可选。

• cf23603: Complete Linux x64 managed browser support: add the pinned Chrome for Testing 146 entry so auto resolves on Linux instead of failing with a missing catalog entry, discover system browsers from their Linux install paths instead of falling through to the Windows branch, and extract zip artifacts with unzip where GNU tar cannot read them.

  完成 Linux x64 managed browser 支持：加入 pinned Chrome for Testing 146 entry，使 auto 在 Linux 上能正常解析；从 Linux 安装路径发现 system browsers，避免错误落入 Windows branch；在 GNU tar 无法读取 zip artifacts 时使用 unzip 解压。

• cf23603: Add an opt-in local API proxy so existing OpenAI- and Anthropic-compatible clients can route Q&A traffic through visible provider sessions: new daemon routes POST /v1/openai/chat/completions, GET /v1/openai/models, and POST /v1/anthropic/messages, explicit tokenless/<provider> model naming, a new-conversation or continue-conversation mapping chosen during setup or with tokenless api-proxy, and fail-closed rejection of tool, function, and structured-output fields that visible pages cannot honour.

  新增 opt-in local API proxy，让现有 OpenAI- 与 Anthropic-compatible clients 可以通过 visible provider sessions 路由问答流量：提供 POST /v1/openai/chat/completions、GET /v1/openai/models 与 POST /v1/anthropic/messages daemon routes，明确使用 tokenless/<provider> model naming，并支持在 setup 或 tokenless api-proxy 中选择 new-conversation 或 continue-conversation mapping；visible pages 无法兑现的 tool、function 与 structured-output fields 会 fail closed。

• 1164999: Preserve fenced code and line breaks in visible-provider response.read results so generated programs can be evaluated without code-block toolbar text or whitespace loss.

  在 visible-provider response.read results 中保留 fenced code 与换行，使生成的程序可以在不混入 code-block toolbar text、也不丢失 whitespace 的情况下被评估。

• 500eeb4: Align the local API proxy with standard stateless Chat Completions and Responses continuation semantics: Chat Completions and Anthropic requests always start fresh provider chats, while a valid Responses previous_response_id resumes its mapped browser conversation and sends only the current input delta.

  让 local API proxy 对齐标准 stateless Chat Completions 与 Responses continuation semantics：Chat Completions 和 Anthropic requests 始终启动新的 provider chats；有效的 Responses previous_response_id 会恢复映射的 browser conversation，并只发送当前 input delta。

• 610136f: Prevent resident managed browsers from accumulating diagnostic chrome://version tabs and released provider about:blank pages across repeated test or worker detach cycles. Real provider, protected, borrowed, and control-plane pages remain open.

  防止驻留的托管浏览器在重复测试或 worker detach 后不断累积诊断用 chrome://version tab 与已释放的 provider about:blank 页面。真实 provider、protected、borrowed 与 control-plane 页面仍保持打开。

• a3c2d40: Harden the local dashboard's daemon connection with a shared typed contract, profile-scoped provider-readiness state, semantic dashboard operations, and immediate secure 404 responses for missing UI assets.

  通过 shared typed contract、profile-scoped provider-readiness state、semantic dashboard operations，以及对缺失 UI assets 立即返回的 secure 404 responses，强化 local dashboard 的 daemon connection。

• 0981a5e: Align the Dashboard header metrics on one line, report idle runtime activity accurately, and add a finished-job count.

  将 Dashboard header 指标对齐为单行，准确显示空闲 runtime 状态，并增加已结束任务计数。

• 293e258: Bind newly created and setup-managed profiles to the exact browser executable that created them, while retaining the existing protection against opening a profile with an older browser version.

  将新建和通过 setup 管理的 managed profile 绑定到创建它的确切 browser executable，同时保留现有的旧版本浏览器保护。

• 4ab236f: Clarify the installed Codex coordination guidance and paired documentation for internal sub-agents, sidebar-visible separate tasks, and concise coordinator reporting.

  补充 Codex coordination guidance 与配套文档，明确 internal sub-agent、侧边栏可见的 separate task，以及简洁的 coordinator 报告规则。

• 86d7b64: Classify uncertain visible prompt-submission failures as unknown-window provider rate limits so automatic routing can fall back before any provider submission occurs.

  将不确定的可见 prompt 提交失败归类为未知窗口的 provider rate limit，使自动路由可以在 provider 尚未收到提交时安全 fallback。

• d95a211: Close every released Tokenless API-owned provider tab after its idle timeout and let the Dashboard reopen an empty profile browser on demand.

  在空闲超时后关闭所有已释放且由 Tokenless API 拥有的 provider tabs，并允许 Dashboard 按需重新打开没有 tab 的 profile 浏览器。

• 536edf9: Bound auto new-conversation provider completion time within the existing request deadline and continue through each untried provider route once after a submitted timeout.

  在现有 request deadline 内限制 auto new-conversation 的单 provider completion 时间，并在 submitted timeout 后依次尝试每条尚未使用的 provider route 一次。

• 657872b: Open the Tokenless Dashboard in the operating system's default browser instead of the selected provider profile browser.

  使用操作系统默认浏览器打开 Tokenless Dashboard，不再使用所选 provider profile 的浏览器。

• d0021fd: Defer providers after a real visible rate-limit observation so auto routing does not immediately select the same exhausted provider again.

  真实可见的 rate-limit observation 出现后暂时 defer 对应 provider，避免 auto routing 立即再次选择同一个已耗尽 provider。

• a7ce915: Fix Dashboard Direct tooltips and links to show each provider's upstream API URL.

  修复 Dashboard Direct tooltips 与 links，使其显示每个 provider 的 upstream API URL。

• 89803cf: Remove AI Badgr and Api.Airforce from the Tokenless API provider catalog.

  从 Tokenless API provider catalog 中移除 AI Badgr 与 Api.Airforce。

• 89803cf: Explain Provider readiness icons and badges with localized hover tooltips, including observed provider-specific subscription tiers.

  为 Provider readiness icons 与 badges 增加 localized hover tooltips，并包含已观察到的 provider-specific subscription tiers 说明。

• 455a289: Mark providers that do not expose subscription tiers and hide the plan indicator for them in the dashboard.

  标记不提供订阅等级的 provider，并在 dashboard 中隐藏其方案指示器。

• 2e24f7d: Fix Arena authentication status detection by inspecting the visible sidebar account surface.

  通过检查可见的侧边栏账户界面，修复 Arena 登录状态检测。

• 89803cf: Increase text sizes throughout the Tokenless Dashboard for easier reading.

  增大 Tokenless Dashboard 的整体文字尺寸，提升可读性。

• b9ed152: Read completed Claude code-block responses, enter Claude prompts through native keyboard input, dismiss its observed transient Not now prompt, clear restored draft attachments and reload the cleared draft before Universal API new conversations, submit through its visible enabled send control even when a non-mutating actionability probe times out, retain native Enter only when no send control is present, preserve sanitized submit-actionability and provider-session diagnostics through the provider-action boundary before safe fallback, route away safely when that control remains disabled, isolate Universal API new conversations on replacement pages while preserving continuations, make the existing one-shot correction explicitly repair invalid structural punctuation and schema-invalid tool arguments, and admit its real-provider-proven single-call strict tool and tool-history scope to tokenless/auto routing.

  读取已完成的 Claude code-block response；通过原生键盘输入 Claude prompt；关闭已观察到的短暂 Not now 弹窗；在 Universal API 新会话开始前清除恢复的旧 draft 附件并 reload 已清空的 draft；即使非变更性 actionability 检查超时，仍通过可见且 enabled 的发送按钮提交，只在完全没有发送控件时保留 Claude 原生 Enter 路径；在安全 fallback 前让净化后的 submit actionability 与 provider-session diagnostics 完整穿过 provider-action 边界，并在控件仍 disabled 时安全路由到下一个 provider；让 Universal API 新会话使用 replacement page 隔离旧 draft，同时保留 continuation page；明确要求现有的一次性 correction 修复无效结构标点与 schema-invalid tool arguments；并将真实 provider 已验证的 single-call strict tool 与 tool-history scope 纳入 tokenless/auto routing。

• c5329ae: Show relative provider and conversation timestamps, readable truncated chat titles, per-provider readiness checks, a clearly limited recent-conversation summary, and direct job-detail links in the local Dashboard overview.

  在 local Dashboard overview 中显示 relative provider 与 conversation timestamps、可读的截断 chat titles、按 provider 的 readiness checks、明确受限的 recent-conversation summary，以及直接的 job-detail links。

• 5e297c8: Show canonical Browser and Direct entry URLs in interactive provider mode tooltips.

  在 interactive provider mode tooltips 中显示 canonical Browser 与 Direct entry URLs。

• c5329ae: Persist provider authentication observations in the shared tokenless.sqlite3 while keeping profile identity and configuration in config.json.

  将 provider authentication observations 持久化到共享的 tokenless.sqlite3，profile identity 与 configuration 继续保存在 config.json。

• 455a289: Give every Tokenless Dashboard page its own pathname instead of a URL hash.

  为每个 Tokenless Dashboard 页面提供独立 pathname，不再使用 URL hash。

• ab8cde5: Require real bootstrap, tool, continuation, and final-result evidence for Browser Harness eligibility while retaining independently verified generic Markdown upload routes.

  要求 Browser Harness eligibility 必须具备真实 bootstrap、tool、continuation 与 final-result evidence，同时保留已独立验证的 generic Markdown upload routes。