@techysy/10router 1.2.0 → 1.3.0

package/app/public/i18n/changelog/en.md

@@ -1,7 +1,62 @@
-# Changelog
-
-User-facing highlights per release. See [CHANGELOG.md](https://github.com/techysy/10router/blob/main/CHANGELOG.md) for the full developer log.
-
+# Changelog
+
+User-facing highlights per release. See [CHANGELOG.md](https://github.com/techysy/10router/blob/main/CHANGELOG.md) for the full developer log.
+
+## v1.3.0 (2026-10-01)
+
+### ✨ New
+
+- **CLI Tools "Model Combo Profiles" (Claude Card, #17)**: The Claude tool card introduces a Profiles section, enabling you to save Opus/Sonnet/Haiku mappings, API Key, and context window settings as named presets. Selecting a preset from the dropdown instantly writes to `~/.claude/settings.json` without requiring an Apply click (parity with CC-Switch). Profiles persist server-side and synchronize across NAS, Tailscale, and tunnel connections.
+- **Quota Visualization Redesign: QuotaToolbar & Window Timeline**: The usage view features a responsive quota toolbar and cycle timeline (`QuotaWindowTimeline`), visualizing rolling reset countdowns across providers. Includes a non-persisted view tab toggle ("Cards" vs. "Quota windows") that always defaults to Cards on fresh page loads. Segmented pack bars align with CreditDaddy visual semantics with absolute nearest-expiry displays.
+- **New Search & Fetch Providers: TinyFish & Keenable (#26)**:
+  - **TinyFish** (Free Tier): Full automated support for Web Search and Web Fetch with zero balance deduction;
+  - **Keenable** (Standard API Key): Realtime/pro search modes and high-fidelity webpage scraping with `live=true`;
+  - Seamlessly integrated across Media Providers and `/v1/search` / `/v1/web/fetch`.
+- **ZCode Free Trial Provider (`zcode-free`)**: Route Start Plan and Trust Build credits (GLM-5.3-Flash) through CreditDaddy desktop gateway, supporting remote/LAN gateway host configuration and detailed architecture documentation.
+- **Qoder Itemized Resource Packs via CreditDaddy Web Session**: Uses synced browser cookies to directly query exact per-pack remaining credits and expiration dates from the web console, avoiding heuristic approximations while validating account ownership.
+- **LongCat-2.5-Preview & International Portal**: Added LongCat-2.5-Preview (1M context, multimodal image/video input, thinking mode toggle) and separate `longcat.ai` provider card.
+- **Desktop In-Container Password Management (`639f77a8`)**: Automatically captures and auto-fills credentials for web apps running in desktop shell containers (WorkBuddy, CodeBuddy), securely encrypted at rest via OS-level safeStorage (DPAPI).
+- **System Theme Adaptive Mode**: Login and dashboard interfaces automatically track OS day/night mode, powered by pre-paint scripts to prevent screen flashing.
+
+### 🐛 Fixes
+
+- **Adaptive reset countdowns**: quota cards and pack bars no longer show a bare reset date — future resets now read as a countdown whose precision follows magnitude (day+hour / hour+minute / minute+second), with the full absolute time still on hover.
+- **5-hour quota view centered on now**: the timeline axis is now the 12 hours around the current moment instead of the calendar day, so "now" always sits mid-track instead of clamped to the right edge by evening.
+- **Usage Input Tokens Normalization**: Fixed an issue where prompt tokens on unfolded Claude usage shapes were overridden by cached token counts, resolving the 1:1 display skew for high-cache providers like `zcode-free`.
+- **Fresh Install Static Model Exposure**: Ensures `noAuth` providers (opencode, mimo-free, zcode-free) expose models in `/v1/models` on brand-new installs with empty connection tables.
+- **OpenAI Responses Content Filter Mapping**: Correctly maps Responses `incomplete_details.reason === "content_filter"` to Chat completion `finish_reason: "content_filter"` instead of `"length"`, preventing client SDKs from assuming context cutoffs.
+- **Schemeless Proxy Normalization & Leak Prevention (#36)**: Accepts proxy URLs without explicit schemes (e.g. `127.0.0.1:7890`, normalized to `http://`). Keeps encrypted strings on parse failures to strictly prevent direct IP leak fallbacks.
+- **Antigravity Upstream Stability**: Removed obsolete `requestType: "agent"` to eliminate spurious 429 errors; fixed empty string tool response 400s; added bidirectional long tool name mapping; reinforced tool schema sanitization with local `$ref` resolution.
+- **Strict Bare Model Resolution (#34)**: Rejects unknown bare model names with a clear 400 error rather than guessing upstream providers and producing misleading 404s.
+- **Remote Provider Management Authorization (#38)**: Restored permission for valid virtual dashboard keys to POST `/api/providers` and `/api/provider-nodes`.
+- **MITM Security Hardening (#31)**: Passed `ROUTER_API_KEY` via stdin in sudo invocation scripts rather than leaking it through process command line arguments.
+
+## v1.2.1 (2026-09-26)
+
+### 🔒 Security
+
+- **[Important] Locked out after turning on the log-in check (#33)**: with no password of your own, switching "Require login" on left only the hidden first-login password (on fnOS it is generated at install time into a file you never see), so every password you tried was rejected. The log-in check can no longer be turned on before you set a password, empty passwords are refused, and there is now a way back in: **"Forgot your password?"** on the login page. On fnOS, just enter a new password in **App Center → 10Router → Settings**; on any install, drop a `reset-password` file into the data folder.
+
+### ✨ New
+
+- **CreditDaddy integration: read-only quota overview for external dashboards**: `GET /api/usage/quotas` returns every quota-capable connection in one call, normalized the same way as the dashboard's Provider Limits, and accepts a dashboard virtual key (`sk-…`). Results are cached for 5 minutes; a forced refresh (`?force=1`) is limited to once per connection every 30 seconds, and simultaneous requests share one upstream call. No credentials are returned — note that account emails are visible to anyone holding a virtual key.
+- **Imported usage now shows an estimated cost**: rows synced in by the 10router-sync plugin or CreditDaddy used to land at $0. New imports are priced on arrival, and existing history is repaired automatically at startup (a source-computed cost is never overwritten; models without a price stay at $0).
+- **The "log-in check is off" banner can be hidden**: Settings → Security gains a switch under *Require login*, shown only while the check is off. Hiding the banner takes a confirmation, and it comes back automatically once you turn log-in back on and off again.
+- **Automatic update checks can be turned off** (Settings → Security): for installs pinned to a version on purpose. When off, 10Router no longer contacts the update server on its own or shows new-version notices — the dashboard, the tray balloon and the CLI launcher all go quiet. "Check now" and the tray's "Check for updates" still work.
+- **Encrypted OAuth import/export on single-auth cards**: the transfer buttons now appear on OAuth-only cards as well — most importantly **MiMo Desktop**, whose signed-in session is the credential you move between machines.
+- **Official Qwen pricing**: new-generation Qwen flagships were priced at the cheap wildcard rate (e.g. qwen3.8-max billed at a quarter of its real price); they now use the official per-model rates.
+
+### 🐛 Fixes
+
+- **[Important] After a restart, remote access stayed down until someone opened the dashboard**: tunnel / Tailscale / MITM auto-resume and the other startup tasks only ran on the first page render — on a NAS used purely through `/v1` they could wait indefinitely. They now start with the server.
+- **Qoder's queue throttle was returned as the model's reply**: a queued request (`10605`) reached the client as `[qoder error 403: …]` text and was logged as a success. It now fails over to the next account and cools the throttled one for the time Qoder asks for (about 30 seconds).
+- **OAuth transfer import was blocked on passwordless dashboards** even from the machine itself; same-machine imports work again.
+- **10router-sync plugin v1.5.0**: follows ZCode's new plan-channel ids (Start Plan traffic had stopped syncing since 09-18), counts mirasim cache in input so dashboards stop showing "input 638, cache 113M", and the mirasim correction script no longer adds the cache twice.
+
+### 🔧 Other
+
+- **Auth header names aligned with 10Router**: `x-10r-cli-token` / `x-10r-password` replace the inherited `x-9r-*` names. The old names are still accepted, so existing CLI launchers, plugins and scripts keep working.
+
 ## v1.2.0 (2026-09-24)
 
 ### ✨ New
@@ -48,235 +103,235 @@ User-facing highlights per release. See [CHANGELOG.md](https://github.com/techys
 - The dashboard's Change Log now only renders released versions — you will not be told about features you do not have yet.
 - Installers and docs no longer point users at an unrelated fork's package name.
 
-## v1.1.3 (2026-09-20)
-
-### ✨ New
-
-- **Qoder CN fully restored + daily credits auto-claim**: qoder-cn OAuth device-code / PAT auth, model catalog and usage tracking are back; daily credit auto-claim (experimental toggle + per-card manual claim) with per-pack resource-package display (individual expiry dates, first-expiry-first-spend); official Qoder icons plus live Qwen price multipliers and off-peak half-price countdown.
-- **Per-model context-window / max-output pins**: a tune icon on every model row (including auto-discovered ones) overrides `contextWindow` / `maxOutput`; leave blank to fall back to defaults. Applies to model listings and usage accounting alike.
-- **Server-side auto-compaction for oversized contexts**: when a request is estimated past 90% of the effective window (threshold selectable 80/90/95%, on by default), older turns are summarized by the same model before dispatch — clients that don't self-compress (ZCode / OpenClaw / custom agents) no longer hit "prompt is too long", with reasoning models handled too; any failure passes the original through untouched. Master toggle lives under Experimental.
-- **Drag-to-reorder provider cards** with persistence; disabled providers auto-sort to the bottom; the model API listing now follows card order.
-- **OpenCode Free anti-abuse fix, enabled by default**; Antigravity image models added (gemini-3-pro-image / imagen-3.0 etc.); 10router-sync plugin v1.5.0 (ZCode plan-channel id adaptation + sync hardening).
-
-### 🛠️ Improvements & Fixes
-
-- **CodeBuddy CN strict-gateway tool-schema compatibility**: tool parameter roots shaped as anyOf/oneOf/allOf, `$ref` or type arrays are now downgraded to a plain object root before forwarding, fixing whole-request `11129` rejections from clients like ZCode (#27).
-- **Quota reset badge no longer disappears**: recurring windows (e.g. MiMo weekly quota) keep showing "resets in N days" even when fully drained.
-- **Qoder quota parsing fix**: daily/campaign credits on free accounts no longer show 0/0.
-- **CodeBuddy intl DeepSeek reasoning_effort mapping fix** (#23).
-- **Stability backports**: 4xx request-level errors no longer trip healthy accounts; successful connection tests clear stale backoff locks; stalled silent streams report in-band error frames instead of looking like normal short replies.
-
-## v1.1.2 (2026-09-18)
-
-### ✨ New
-
-- **Usage dashboard overhaul**: Lifetime stats cards, a GitHub-style activity heatmap (day/week views with hover details), and a node health score table (success rate / latency / speed weighted, expandable per-model rows) on the Usage Details page. The unit-abbreviation toggle now applies globally — overview cards follow the same switch.
-- **Xiaomi Token Plan egress-region matching**: Adding or editing a connection auto-selects the cn / sgp / ams cluster based on your network egress (multi-source probing, fail-open), and the connection test button now routes to the selected cluster. MiMo 429 responses carry friendly guidance, and cooldowns honor the upstream "retry after N seconds" hint.
-- **10router-sync plugin v1.4.0**: New `/10router-sync:status` command — instance status and today's usage without opening the dashboard.
-
-### 🛠️ Improvements & Fixes
-
-- **CodeBuddy 11128 channel-level circuit breaker**: Channel-wide risk-control hits now trip a provider-level cooldown (no more per-account retry bursts amplifying risk), with bilingual guidance in the error response; extreme request sizes are backstopped locally.
-- **Credential auto-refresh fixed (Cline / ClinePass)**: Background refresh no longer fails with 400 — credentials renew automatically before expiry; dead credentials are explicitly flagged as needing re-authorization.
-- **Provider & chart cleanup**: Retired public endpoints removed; OpenCode Free / MiMo Code Free moved to a hidden-by-default "Trial" category; model family aggregation fixed (gpt-6, hy4, qwen3.8 etc. are no longer split); Model Type chart legend moved to the top-right with mobile responsiveness.
-- **Security**: The single-connection API no longer echoes the Xiaomi desktop session credential in plaintext.
-
-## v1.1.1 (2026-09-14)
-
-### ✨ New
-
-- **Earliest Expiry First multi-account routing strategy**: Solves the pain point of repeatedly exhausting a primary account while short-term promotional packages on other accounts expire unused. A dedicated toggle on the provider details page automatically prioritizes valid packages closest to expiration, with zero blocking and zero request latency.
-- **Command Code usage tracking & quota dashboard (#16)**: Full support for 5-hour session rolling windows, 7-day weekly limits, and monthly plan credit allowances/spend. Dynamically detects Go / GOAT / Pro subscription tiers, with countdown timers and balance countdowns.
-- **Bulk enable/disable for custom & passthrough models**: One-click bulk enable or disable for custom model catalogs on both OpenAI-compatible endpoints and built-in providers, plus a fix preventing single model updates from overwriting capability attributes.
-- **Universal OAuth credential transfer with AES-256-GCM encryption**: Export credentials across all OAuth providers into secure password-protected backups (`10router-oauth-secure-v1`), with smart multi-tier deduplication during import.
-- **Official brand SVG icons**: Integrated `@lobehub/icons` with official vector brand icons for 58 providers, supporting automatic light/dark theme adaptation.
-- **CodeBuddy check-in & international daily active session**: Check-in upgraded to continuous background scheduling with persistent daily memo deduplication; added daily active session toggle for CodeBuddy International.
-- **Desktop client improvements**: Full tri-lingual localization for menu bars, new "Go" navigation menu (open URL / return to 10Router / recent URLs manager), and long-term server log archiving by date.
-
-### 🐛 Fixed
-
-- **Quota panel 0 total & sentinel countdown bug**: Fixed logic where zero-credit accounts were rendered as infinite quota `0 / ∞`, and filtered out far-future sentinel timestamps (such as year 9999) that produced 2.9-million-day countdowns.
-- **Quota percentage calculation**: Fixed balance-denominated quotas (e.g. Command Code $9.98) being misread as raw percentages displaying `10% 🔴`.
-- **CodeBuddy quota display direction**: Inverted credit pack displays to "Remaining / Total" countdown, aligning numbers with progress bars.
-- **CodeBuddy International connection testing**: Added missing IDE transport headers and auth verification handling, resolving "Provider test not supported".
-- **Xiaomi MiMo Desktop Preview model errors**: Clarified errors when calling preview models without desktop login, eliminating misleading rate-limit cooldowns.
-- **Instant refresh for model disabled states**: Synchronized model availability state immediately upon connection changes, allowing disabled models to be tested directly.
-- **Extensive i18n cleanup**: Full English and Traditional Chinese localization across the Endpoint page, provider notices, and API key buttons.
-
-## v1.1.0 (2026-09-11)
-
-> **Version note**: v1.0.9 was skipped and never released — its planned content (Xiaomi MiMo desktop support, etc.) was folded into this release, which grew to minor-level scope; the version number moved straight to v1.1.0. No v1.0.9 tag or artifacts exist.
-
-### ✨ New
-
-- **Xiaomi MiMo Desktop support (one provider, two ways to sign in)**: `xiaomi-mimo` now accepts both an sk- API key (cloud API) and a Xiaomi MiMo desktop-app account — the desktop-exclusive `mimo-x-pro-preview` / `mimo-x-flash-preview` only accept the account cookie, so an API key alone cannot reach them. Both credentials share one provider card; enable whichever you need.
-- **Xiaomi MiMo authorization-code sign-in**: after the official authorize page opens in your browser, you can **paste the code shown on that page** straight into the dialog to finish signing in — even if the local callback never connected, or already timed out. The credential is decrypted server-side only and never passes through the browser.
-- **CodeBuddy international catalog completed**: every entry was re-checked against the live gateway with real requests, adding five models that were served but missing (GLM-5.1 / GLM-5v-Turbo / MiniMax-M3 / Kimi-K2.7 / DeepSeek-V4.1-Flash) plus `GPT 6.0 Astra`.
-- **opencode-go catalog aligned with the official public list (18 models), with a declared endpoint per model**: some models used to fail outright because a missing endpoint declaration sent them down the wrong path; each one now declares its endpoint from the official table.
-- **New `gpt-image-2.5` image-model family for Codex / OpenAI**: `gpt-image-2.5` (incl. `-flare` / `-sunburst`), `gpt-image-2`, `gpt-image-1.5` — text-to-image, editing, and multi-image references.
-- **Credit-multiplier badge supports limited-time free promos**: the badge shows green `free` during the promo and **automatically** returns to the real multiplier when it ends — no one has to remember to change it back.
-
-### 🐛 Fixed
-
-- **Windows desktop upgrades no longer silently discard old data**: the desktop data directory is also Electron's own config directory, so Chromium fills it with files first and makes it look "non-empty", permanently skipping the old-data migration. The check is now "does the directory already contain our own database".
-- **Several real Xiaomi MiMo authorization-code bugs**: the decrypted payload layout was reversed (the actual cause of `decrypt_failed`), the platform requires `app=MiMo`, the callback path is now a random string, and the result is handed back with a 302; pasting a code also keeps working after the callback listener times out.
-- **Kiro requests no longer carry a top-level `systemPrompt`** (kiro.dev answers it with an immediate 400 and no retry); endpoint order now prefers the Amazon surface.
-- **Claude tool `type` is now decided per provider**: this fixes MiniMax's Claude endpoint (rejected tools without `type`) without breaking DeepSeek's Claude endpoint (which only accepts its own `type` values).
-- **Codex strips `\p{...}` regexes it cannot parse from tool schemas**: previously a single property escape in a tool parameter `pattern` failed the whole request with a 400, identically for every account.
-- **`better-sqlite3` is skipped on Node 24**: its native addon crashes the process on load there, which used to take the whole server down.
-- **AMD Token Factory catalog and capabilities rebuilt from live measurements**: missing chat models were added and a "text-only" misclassification was fixed (images were being silently dropped).
-- **CodeBuddy CN's DeepSeek-V4.1-Flash output ceiling lowered to the server-published 128K**: asking for more never errors, it just makes over-limit requests more likely to hit a 400.
-- **One provider no longer stores two contradictory disabled-model records**: this fixes models that appeared not to be disabled.
-
-## v1.0.8 (2026-09-09)
-
-### ✨ New
-- **AMD Token Factory provider**: free shared OpenAI-compatible endpoints from AMD Radeon Cloud (`DeepSeek-V4-Flash` 1M context + `Qwen3.8-Flash-Next` 262K context); thinking via reasoning_effort with per-model level pickers; same shape as NVIDIA NIM
-- **Antigravity quota host fix**: quota summary now queries the correct daily host (matching the native IDE), fixing systematically stale numbers; percentage reads from the backend value instead of re-deriving in the frontend
-- **CodeBuddy catalogs aligned to the server (CN + intl)**: the international catalog is re-aligned (adds the Hy4-Preview / Hy3 free tier, GPT-5.6 Sol/Terra/Luna, GLM-5.3, Kimi-K3; drops retired models); the CN catalog now matches too — `DeepSeek-V4.1-Flash` supersedes `DeepSeek-V4-Flash` and the duplicate `Kimi-K3 (1)` slot is gone; every model on both gateways now carries the published credit multiplier
-- **Credit-multiplier badges on the model list**: free-quota models show a green `free` chip, the rest show their modifier as `0.79x` (hover for an explanation); providers with no credit system show nothing
-
-### 🔒 Security
-- **API key HMAC secret hardening**: built-in fallback secret now triggers a production startup warning; experimental **Key secret rotation** (off by default) auto-generates a per-install secret; `generateKeyId` switched from `Math.random` to `crypto.randomBytes`
-- **Dangerous-action confirm dialogs**: "Require API key" off, key-rotation enable/disable, and "Rotate all" now require explicit confirmation; rotation is double-submit protected (server 409 + client guard)
-
-### 🐛 Fixed
-- **Non-streaming requests defaulted to SSE (issue #4)**: omitting `stream` now correctly defaults to JSON (OpenAI/Anthropic spec); previously returned `text/event-stream` with trailing `data: [DONE]` that broke strict JSON clients (OpenAI SDK, WorkBuddy, curl)
-- **Key rotation confirm dialog stuck open / repeated rotation**: confirm modal now auto-closes before executing; repeated rotations no longer stack "(rotated)" suffixes
-- **MCP SSE idle disconnect**: long-lived MCP SSE connections now send a 25s comment heartbeat, preventing NAT/firewall silent drops
-- **CodeBuddy CN whitelist false-positive (PR #5)**: Claude Code's own system prompt no longer matches the agent-identity whitelist (was triggering 11128)
-- **GLM-4.6V-Flash missing vision (PR #5)**: free vision model now registered with correct capabilities
-- **DeepSeek effort "max" rejected by SenseNova (PR #5)**: clamped to "xhigh" for cross-provider compatibility
-- **Fetched/imported models now default to disabled (enable on demand)**: bulk imports via "Import from /models" and "Fetch Qoder Models" no longer switch everything on at once — they land in the "Disabled models" area for you to enable as needed, and `/v1/models` now only serves enabled models (a disabled model was previously still listed). Single manual adds stay enabled
-- **CodeBuddy CN listing intl-only models (fixed)**: the CN catalog had picked up the GPT/Gemini family that only the international gateway serves, which failed with "model service info not found" when selected; removed from CN (kept in intl) with a test guard
-- **DeepSeek-V4.1-Flash output ceiling under-reported (fixed)**: its capabilities capped output at 50K (a leftover from the pre-rename id), which truncated client requests; corrected to the model card's 1M context / 384K output / image input / switchable thinking, and the model's two alias ids now carry the same caps
-
-## v1.0.7 (2026-09-07)
-
-### ✨ New
-- **CodeBuddy CN built-in "Total Points" quota row**: the CodeBuddy CN card now leads with a Total Points row summing the live balance of every pack (monthly refills + bonus packs) — no more adding them up yourself; it auto-hides under the "Only with balance" filter once drained; "Bonus Pack" row names are now localized
-- **ZCode usage sync plugin (10router-sync)**: import ZCode's local model-usage ledger into 10Router in one step — providers pointed at 10Router are auto-excluded to prevent double counting; re-runs are idempotent via dedup; imported usage shows up in the usage page and the details tab
-- **Upstream v0.5.69 selective port**: ① Google-family multi-account background refresh → serial + tiered jitter to dodge anti-abuse risk control; ② Anthropic-compatible nodes on real Claude auto-add context-management beta headers (no more silent model-swap); ③ opencode-go stable sessions (free pool no longer trips risk control); ④ Responses parallel tool-call fix (no longer merges N calls into one); ⑤ Claude Fable weekly-quota tracking; ⑥ codex adds gpt-6-astra etc.; ⑦ codebuddy-cn model catalog aligned to server contract
-- **qoder catalog refresh + image passthrough**: model list aligned to server, base64 image direct send; dashboard Antigravity quota grouped by family (per-connection isolation, better than upstream)
-- **Desktop tray update path + leaner menu**: desktop update banner now points to GitHub Releases; tray menu merges check-update / about / start-stop into one compact flow
-- **copilot switched to VS Code extension guide**: no more MITM intercept — three-step extension config (engine layer retained, manually revertible)
-- **Tray icon monochrome**: macOS template icon + Windows dark/light theme black-white adaptive
-
-### 🐛 Fixed
-- **`/responses` root-path auth gap (security)**: prefix added, path now enforces API-key check
-- **Foreign `server_tool_use` poisoning history after Claude combo fallback (400)**: validate `srvtoolu_` prefix and drop foreign tool blocks
-- **MCP deferred tool breaking cache anchors (400)**: anchor now pinned to the last cacheable tool
-- **gemini schema tuple validation 400**: `prefixItems` conversion + array missing `items` placeholder
-- **antigravity system-prompt competitor brand-clean generalization**: OpenCode naming no longer triggers 429
-- **Free-model background-refresh noise + connection-test hang**: de-noised + 15s timeout
-
-## v1.0.6 (2026-09-05)
-
-### ✨ New
-
-- **CodeBuddy CN auto daily check-in** (experimental, off by default): with "CodeBuddy CN auto daily check-in" enabled under Settings → Experimental, the Import / Export buttons on the CodeBuddy CN page are replaced by an automatic daily check-in (each account checks in at a random 00:00–06:00 local time to renew free quota; failures never interrupt the service and a 401 triggers one token refresh + retry) plus a manual "Check in now" button
-- **Antigravity quota aligned with the official site (5h + weekly dual window)**: Antigravity quotas on the usage page now use the official retrieveUserQuotaSummary RPC as the primary source, showing exactly **4 cards** per account (Gemini Models / Claude and GPT models × {5-hour, weekly}), normalized to a 0–100 scale; falls back to the old per-model parse when the RPC is unavailable
-- **OpenCode Free catalog aligned with the official free list**: adds Big Pickle / MiMo-V2.5 / Ling 3.0 Flash Fin / Nemotron 3 Ultra / Nemotron 3.5 Lightning; deliberately excludes the DeepSeek / Laguna free tiers the official docs have dropped (avoids accidental paid calls)
-
-### 🐛 Fixes
-- **Antigravity `ag/gemini-3.8-flash-*` 404**: model addressing now uses per-tier tiered entities and the IDE fingerprint is bumped to 2.11.0 (bare ids 404 with "Requested entity was not found"); adds the un-tiered `ag/gemini-3.8-flash` (routed to the medium tier)
-- **"Only with balance" misses newly-checked-in quota packs**: fixed a case where CodeBuddy CN's daily full packs were hidden after older packs were renumbered and the persistent hide never cleared — the hidden set is now recomputed live so any pack with remaining balance (including fresh full packs) always shows
-- **Friendly message for free-model 429 rate limits**: when a free-tier model (`oc/*-free`, `contributor-free`, APInex `free/`-prefixed, etc.) hits an upstream rate limit and returns 429, the raw English `rate_limit_exceeded` is no longer passed through — instead a friendly Chinese message is returned showing an estimated wait ("try again in ~N sec/min/hr") when the upstream supplied a reset time, otherwise advising to retry later or switch to a paid tier. Paid models and multi-account fallback are untouched
-
-## v1.0.5 (2026-09-03)
-
-### ✨ New
-
-- **Desktop tray app (Windows / macOS)**: an install-and-go desktop app — system tray menu (open console / start / restart / stop / launch at login) plus an embedded window; closing the window minimizes to the tray. Shares data, keys and port with the npm CLI (the two forms run exclusively). Windows ships an NSIS installer (silent `/S` supported) and a portable exe; macOS ships separate Intel + Apple Silicon dmgs
-- **CLI / desktop UI localization**: the npm CLI and the desktop tray now follow the system language (Simplified Chinese / Traditional Chinese / English); override with the `TENROUTER_LANG` env var (`zh-CN` / `zh-TW` / `en`)
-- **Self-serve custom providers + new Skills-page skill**: register a custom OpenAI/Anthropic-compatible endpoint (baseUrl + upstream key + models) at runtime with a dashboard LLM key — no source change, no repackage. The Dashboard **Skills** page gains a **10router-add-provider** card and now links to the `main` branch (previously `master`, which 404'd)
-- **CodeBuddy CN account JSON bulk import / export** (experimental, off by default): after enabling "CodeBuddy CN OAuth import / export" under Settings → Providers, the CodeBuddy CN detail page shows Import / Export buttons to bulk-import or export account credentials in the third-party (wb) JSON format (import dedupes by identity and skips non-CodeBuddy issuers). Import / export now require a dashboard-password confirmation (prevents anonymous token export in no-login mode)
-- **Community welfare providers show by default**: GoRouter / TaBiAI now appear by default (no need to toggle them on), consistent across the provider list / Profile toggle / usage topology
-- **Community welfare providers sort into one group**: in the Free Tier list, community providers (GoRouter / TaBiAI) now cluster into one adjacent block per rank group instead of interleaving with regular freeTier providers by priority/name
-- **New Agnes AI dual-site providers**: international **Agnes AI** (com) + **Agnes AI (CN)** — each with Agnes 2.5 Flash / 2.5 Pro text models (512K / 1M context, vision + reasoning), plus Agnes Image 2.x Flash image models (standard images/generations endpoint; image-to-image / editing)
-- **New "Experimental" settings group**: Profile gains a dedicated Experimental card gathering the developer-oriented default-off toggles (Fetch models from GitHub JSON + CodeBuddy CN import/export) for easy future expansion
-- **New provider APInex (apinex.bond)**: prepaid USD-credit third-party gateway (OpenAI-compatible) with 18 models (13 paid + 5 `free/`-prefixed free models); provider pages show a copyable invite-code chip (APInex: `SLEWP68C`)
-- **Upstream v0.5.65 catalog sync + new search providers**: adds GLM-5.3-Flash(Vision), GLM-4.6V, DeepSeek-V4-Flash-Vision-Exp, Grok-4.6, Claude-Fable-5.1 and more; new search providers Ollama-Search (reuses your local Ollama key) and Xquik (X/Twitter search)
-
-### 🐛 Fixes
-- **Usage table corruption after Cost/Token switch**: toggling the display mode and sorting no longer reverts some rows to currency values; Traditional Chinese usage-table headers are now fully translated
-- **Usage table sorting semantics**: token / cost columns now sort model groups by their totals, matching the header arrow direction
-
-## v1.0.4 (2026-09-01)
-
-### ✨ New
-- **3 new providers**: **TokenBom** (decentralized token marketplace — idle API keys earn credits, credits call many models; 79-model online catalog), **GoRouter** (free gateway, no recharge entry), **TaBiAI** (free gateway, no recharge entry)
-- **Usage history import**: import usage data from 9Router backups (SQLite files) — merged into stats without touching config
-- **Notification overhaul**: global toasts moved to top-center; browser-native alert() dialogs replaced with friendly notifications
-- **Community welfare providers**: GoRouter / TaBiAI are hidden by default with a "community" badge; toggle them on in **Settings → Providers → Show community welfare providers**
-
-### 🐛 Fixes
-- **Provider ordering fixed**: connected providers first, disabled sink to bottom — no longer scrambled by priority
-- **JSON catalog models invisible after enabling**: stale disable records are now cleared ("enabled but not listed" resolved)
-- **B.AI / CodeBuddy CN model catalogs completed**: missing models that caused "not found" when switching are added
-- **Friendly maintenance hints on connection tests**: endpoint down / Cloudflare-blocked shows a maintenance note instead of a misleading "Invalid API key"
-- **Account filter reminder**: quota-pack account filter persists; an amber reminder bar shows when a non-default filter is active
-- **CodeBuddy CN DeepSeek 11150**: DeepSeek-series calls no longer fail with 400 on the reasoning-effort param (auto/off); coding agents (dsh, etc.) work normally
-- **CodeBuddy empty tool name (11133 / unknown tool)**: empty `function.name` in streaming tool calls is normalized, so standard clients no longer mis-detect the tool name
-- **Topology still shows hidden community providers**: the usage topology now honors the "Show community providers" toggle — community welfare sites (GoRouter/TaBiAI) are hidden when it's off
-- **Skills page i18n + Chinese links**: Skills page text is now localized; in Chinese (zh-CN) the links point to the Chinese skill files
-
-## v1.0.3 (2026-08-30)
-
-### ✨ New
-- **4 new providers**: **LongCat** (Meituan), **SenseNova** (SenseTime, free beta), **Dots** (Xiaohongshu Dots Studio, free beta), **B.AI** (aggregator — one key for GPT / Claude / Gemini / DeepSeek / GLM / Kimi / Qwen and more)
-- **Model JSON catalogs for custom providers**: custom nodes can pull an online model list, toggle models on/off individually, and manage them in bulk
-- **fpk update check goes straight to Releases**: fnOS installs jump directly to the matching release download
-
-### 🔒 Security
-- **Progressive login rate-limit**: 5 failed password attempts → 30s / 2m / 10m / 30m lockout
-- **Reject placeholder JWT keys**: copying the public `.env.example` key is ignored; a random key is generated instead
-- **Fix npm package leaking build-machine secrets**: no more keys / machine IDs / data snapshots in the build artifact
-
-### ⚠️ Notes
-- npm package renamed to **`@techysy/10router`** (the old `10router` is an unrelated fork). If you installed `10router-cli`, switch to the new package; data directory unchanged.
-
-## v1.0.2 (2026-08-29)
-
-### 🐛 Fixes
-- **Update check no longer points at a third-party package** (affected 1.0.1): version check, update command, and sidebar install command all target the correct package
-- **postinstall no longer aborts install**: npm install no longer fails from the warm-up script on WSL paths
-
-### ⚙️ Engineering
-- **Test CI added**: tests + regression gate run automatically on push / PR
-
-> ⚠️ **1.0.1 users should upgrade**: its built-in "check for updates" points to an unrelated third-party package.
-
-## v1.0.1 (unreleased; delivered with v1.0.2)
-
-> 1.0.1 was never published as a release (only the npm `10router-cli@1.0.1` briefly existed). The following reached Docker / fpk / standalone users with **v1.0.2**.
-
-### 🔒 Security
-- **4 MITM security fixes**: upstream TLS cert validation restored, root CA private key locked to 0600, no longer blindly killing the process on port 443, auto-cleanup of leftover hosts entries (MITM is off by default)
-
-### ✨ New
-- **Disabled providers sink to the bottom** via a settings toggle
-- **Collapsible desktop sidebar**
-- **OpenCode Go quota usage**
-- **Gitee mirror fallback for model catalogs** (faster in CN)
-
-### 🐛 Fixes
-- **/v1/models no longer returns orphan custom models**
-- **Custom node prefix uniqueness check**
-- **Fix CodeBuddy executor dropping the Agent system prompt**
-
-### ⚙️ Engineering
-- **New npm distribution channel**: `npm i -g 10router-cli`
-- **Generic "fetch models from GitHub JSON" capability** (Fetch Models)
-
-## v1.0.0 (2026-08-26)
-
-### ⚠️ Notes
-1. **Data directory renamed**: `~/.9router/` → `~/.10router/` (Windows: `%APPDATA%\10router`), auto-migrated on first start
-2. **SAML entityID changed**: default issuer is now `urn:10router:sp`; re-register in your IdP
-3. **MITM CA renamed**: re-trust the new CA
-
-### ✨ New
-- **Rebrand**: 9Router → 10Router
-- **i18n multi-region currency**: en / pt-BR / pt-PT / es / de
-- **Multi-platform distribution**: Docker (amd64 / arm64), fnOS fpk (x86 / arm × url / iframe), Standalone
+## v1.1.3 (2026-09-20)
+
+### ✨ New
+
+- **Qoder CN fully restored + daily credits auto-claim**: qoder-cn OAuth device-code / PAT auth, model catalog and usage tracking are back; daily credit auto-claim (experimental toggle + per-card manual claim) with per-pack resource-package display (individual expiry dates, first-expiry-first-spend); official Qoder icons plus live Qwen price multipliers and off-peak half-price countdown.
+- **Per-model context-window / max-output pins**: a tune icon on every model row (including auto-discovered ones) overrides `contextWindow` / `maxOutput`; leave blank to fall back to defaults. Applies to model listings and usage accounting alike.
+- **Server-side auto-compaction for oversized contexts**: when a request is estimated past 90% of the effective window (threshold selectable 80/90/95%, on by default), older turns are summarized by the same model before dispatch — clients that don't self-compress (ZCode / OpenClaw / custom agents) no longer hit "prompt is too long", with reasoning models handled too; any failure passes the original through untouched. Master toggle lives under Experimental.
+- **Drag-to-reorder provider cards** with persistence; disabled providers auto-sort to the bottom; the model API listing now follows card order.
+- **OpenCode Free anti-abuse fix, enabled by default**; Antigravity image models added (gemini-3-pro-image / imagen-3.0 etc.); 10router-sync plugin v1.5.0 (ZCode plan-channel id adaptation + sync hardening).
+
+### 🛠️ Improvements & Fixes
+
+- **CodeBuddy CN strict-gateway tool-schema compatibility**: tool parameter roots shaped as anyOf/oneOf/allOf, `$ref` or type arrays are now downgraded to a plain object root before forwarding, fixing whole-request `11129` rejections from clients like ZCode (#27).
+- **Quota reset badge no longer disappears**: recurring windows (e.g. MiMo weekly quota) keep showing "resets in N days" even when fully drained.
+- **Qoder quota parsing fix**: daily/campaign credits on free accounts no longer show 0/0.
+- **CodeBuddy intl DeepSeek reasoning_effort mapping fix** (#23).
+- **Stability backports**: 4xx request-level errors no longer trip healthy accounts; successful connection tests clear stale backoff locks; stalled silent streams report in-band error frames instead of looking like normal short replies.
+
+## v1.1.2 (2026-09-18)
+
+### ✨ New
+
+- **Usage dashboard overhaul**: Lifetime stats cards, a GitHub-style activity heatmap (day/week views with hover details), and a node health score table (success rate / latency / speed weighted, expandable per-model rows) on the Usage Details page. The unit-abbreviation toggle now applies globally — overview cards follow the same switch.
+- **Xiaomi Token Plan egress-region matching**: Adding or editing a connection auto-selects the cn / sgp / ams cluster based on your network egress (multi-source probing, fail-open), and the connection test button now routes to the selected cluster. MiMo 429 responses carry friendly guidance, and cooldowns honor the upstream "retry after N seconds" hint.
+- **10router-sync plugin v1.4.0**: New `/10router-sync:status` command — instance status and today's usage without opening the dashboard.
+
+### 🛠️ Improvements & Fixes
+
+- **CodeBuddy 11128 channel-level circuit breaker**: Channel-wide risk-control hits now trip a provider-level cooldown (no more per-account retry bursts amplifying risk), with bilingual guidance in the error response; extreme request sizes are backstopped locally.
+- **Credential auto-refresh fixed (Cline / ClinePass)**: Background refresh no longer fails with 400 — credentials renew automatically before expiry; dead credentials are explicitly flagged as needing re-authorization.
+- **Provider & chart cleanup**: Retired public endpoints removed; OpenCode Free / MiMo Code Free moved to a hidden-by-default "Trial" category; model family aggregation fixed (gpt-6, hy4, qwen3.8 etc. are no longer split); Model Type chart legend moved to the top-right with mobile responsiveness.
+- **Security**: The single-connection API no longer echoes the Xiaomi desktop session credential in plaintext.
+
+## v1.1.1 (2026-09-14)
+
+### ✨ New
+
+- **Earliest Expiry First multi-account routing strategy**: Solves the pain point of repeatedly exhausting a primary account while short-term promotional packages on other accounts expire unused. A dedicated toggle on the provider details page automatically prioritizes valid packages closest to expiration, with zero blocking and zero request latency.
+- **Command Code usage tracking & quota dashboard (#16)**: Full support for 5-hour session rolling windows, 7-day weekly limits, and monthly plan credit allowances/spend. Dynamically detects Go / GOAT / Pro subscription tiers, with countdown timers and balance countdowns.
+- **Bulk enable/disable for custom & passthrough models**: One-click bulk enable or disable for custom model catalogs on both OpenAI-compatible endpoints and built-in providers, plus a fix preventing single model updates from overwriting capability attributes.
+- **Universal OAuth credential transfer with AES-256-GCM encryption**: Export credentials across all OAuth providers into secure password-protected backups (`10router-oauth-secure-v1`), with smart multi-tier deduplication during import.
+- **Official brand SVG icons**: Integrated `@lobehub/icons` with official vector brand icons for 58 providers, supporting automatic light/dark theme adaptation.
+- **CodeBuddy check-in & international daily active session**: Check-in upgraded to continuous background scheduling with persistent daily memo deduplication; added daily active session toggle for CodeBuddy International.
+- **Desktop client improvements**: Full tri-lingual localization for menu bars, new "Go" navigation menu (open URL / return to 10Router / recent URLs manager), and long-term server log archiving by date.
+
+### 🐛 Fixed
+
+- **Quota panel 0 total & sentinel countdown bug**: Fixed logic where zero-credit accounts were rendered as infinite quota `0 / ∞`, and filtered out far-future sentinel timestamps (such as year 9999) that produced 2.9-million-day countdowns.
+- **Quota percentage calculation**: Fixed balance-denominated quotas (e.g. Command Code $9.98) being misread as raw percentages displaying `10% 🔴`.
+- **CodeBuddy quota display direction**: Inverted credit pack displays to "Remaining / Total" countdown, aligning numbers with progress bars.
+- **CodeBuddy International connection testing**: Added missing IDE transport headers and auth verification handling, resolving "Provider test not supported".
+- **Xiaomi MiMo Desktop Preview model errors**: Clarified errors when calling preview models without desktop login, eliminating misleading rate-limit cooldowns.
+- **Instant refresh for model disabled states**: Synchronized model availability state immediately upon connection changes, allowing disabled models to be tested directly.
+- **Extensive i18n cleanup**: Full English and Traditional Chinese localization across the Endpoint page, provider notices, and API key buttons.
+
+## v1.1.0 (2026-09-11)
+
+> **Version note**: v1.0.9 was skipped and never released — its planned content (Xiaomi MiMo desktop support, etc.) was folded into this release, which grew to minor-level scope; the version number moved straight to v1.1.0. No v1.0.9 tag or artifacts exist.
+
+### ✨ New
+
+- **Xiaomi MiMo Desktop support (one provider, two ways to sign in)**: `xiaomi-mimo` now accepts both an sk- API key (cloud API) and a Xiaomi MiMo desktop-app account — the desktop-exclusive `mimo-x-pro-preview` / `mimo-x-flash-preview` only accept the account cookie, so an API key alone cannot reach them. Both credentials share one provider card; enable whichever you need.
+- **Xiaomi MiMo authorization-code sign-in**: after the official authorize page opens in your browser, you can **paste the code shown on that page** straight into the dialog to finish signing in — even if the local callback never connected, or already timed out. The credential is decrypted server-side only and never passes through the browser.
+- **CodeBuddy international catalog completed**: every entry was re-checked against the live gateway with real requests, adding five models that were served but missing (GLM-5.1 / GLM-5v-Turbo / MiniMax-M3 / Kimi-K2.7 / DeepSeek-V4.1-Flash) plus `GPT 6.0 Astra`.
+- **opencode-go catalog aligned with the official public list (18 models), with a declared endpoint per model**: some models used to fail outright because a missing endpoint declaration sent them down the wrong path; each one now declares its endpoint from the official table.
+- **New `gpt-image-2.5` image-model family for Codex / OpenAI**: `gpt-image-2.5` (incl. `-flare` / `-sunburst`), `gpt-image-2`, `gpt-image-1.5` — text-to-image, editing, and multi-image references.
+- **Credit-multiplier badge supports limited-time free promos**: the badge shows green `free` during the promo and **automatically** returns to the real multiplier when it ends — no one has to remember to change it back.
+
+### 🐛 Fixed
+
+- **Windows desktop upgrades no longer silently discard old data**: the desktop data directory is also Electron's own config directory, so Chromium fills it with files first and makes it look "non-empty", permanently skipping the old-data migration. The check is now "does the directory already contain our own database".
+- **Several real Xiaomi MiMo authorization-code bugs**: the decrypted payload layout was reversed (the actual cause of `decrypt_failed`), the platform requires `app=MiMo`, the callback path is now a random string, and the result is handed back with a 302; pasting a code also keeps working after the callback listener times out.
+- **Kiro requests no longer carry a top-level `systemPrompt`** (kiro.dev answers it with an immediate 400 and no retry); endpoint order now prefers the Amazon surface.
+- **Claude tool `type` is now decided per provider**: this fixes MiniMax's Claude endpoint (rejected tools without `type`) without breaking DeepSeek's Claude endpoint (which only accepts its own `type` values).
+- **Codex strips `\p{...}` regexes it cannot parse from tool schemas**: previously a single property escape in a tool parameter `pattern` failed the whole request with a 400, identically for every account.
+- **`better-sqlite3` is skipped on Node 24**: its native addon crashes the process on load there, which used to take the whole server down.
+- **AMD Token Factory catalog and capabilities rebuilt from live measurements**: missing chat models were added and a "text-only" misclassification was fixed (images were being silently dropped).
+- **CodeBuddy CN's DeepSeek-V4.1-Flash output ceiling lowered to the server-published 128K**: asking for more never errors, it just makes over-limit requests more likely to hit a 400.
+- **One provider no longer stores two contradictory disabled-model records**: this fixes models that appeared not to be disabled.
+
+## v1.0.8 (2026-09-09)
+
+### ✨ New
+- **AMD Token Factory provider**: free shared OpenAI-compatible endpoints from AMD Radeon Cloud (`DeepSeek-V4-Flash` 1M context + `Qwen3.8-Flash-Next` 262K context); thinking via reasoning_effort with per-model level pickers; same shape as NVIDIA NIM
+- **Antigravity quota host fix**: quota summary now queries the correct daily host (matching the native IDE), fixing systematically stale numbers; percentage reads from the backend value instead of re-deriving in the frontend
+- **CodeBuddy catalogs aligned to the server (CN + intl)**: the international catalog is re-aligned (adds the Hy4-Preview / Hy3 free tier, GPT-5.6 Sol/Terra/Luna, GLM-5.3, Kimi-K3; drops retired models); the CN catalog now matches too — `DeepSeek-V4.1-Flash` supersedes `DeepSeek-V4-Flash` and the duplicate `Kimi-K3 (1)` slot is gone; every model on both gateways now carries the published credit multiplier
+- **Credit-multiplier badges on the model list**: free-quota models show a green `free` chip, the rest show their modifier as `0.79x` (hover for an explanation); providers with no credit system show nothing
+
+### 🔒 Security
+- **API key HMAC secret hardening**: built-in fallback secret now triggers a production startup warning; experimental **Key secret rotation** (off by default) auto-generates a per-install secret; `generateKeyId` switched from `Math.random` to `crypto.randomBytes`
+- **Dangerous-action confirm dialogs**: "Require API key" off, key-rotation enable/disable, and "Rotate all" now require explicit confirmation; rotation is double-submit protected (server 409 + client guard)
+
+### 🐛 Fixed
+- **Non-streaming requests defaulted to SSE (issue #4)**: omitting `stream` now correctly defaults to JSON (OpenAI/Anthropic spec); previously returned `text/event-stream` with trailing `data: [DONE]` that broke strict JSON clients (OpenAI SDK, WorkBuddy, curl)
+- **Key rotation confirm dialog stuck open / repeated rotation**: confirm modal now auto-closes before executing; repeated rotations no longer stack "(rotated)" suffixes
+- **MCP SSE idle disconnect**: long-lived MCP SSE connections now send a 25s comment heartbeat, preventing NAT/firewall silent drops
+- **CodeBuddy CN whitelist false-positive (PR #5)**: Claude Code's own system prompt no longer matches the agent-identity whitelist (was triggering 11128)
+- **GLM-4.6V-Flash missing vision (PR #5)**: free vision model now registered with correct capabilities
+- **DeepSeek effort "max" rejected by SenseNova (PR #5)**: clamped to "xhigh" for cross-provider compatibility
+- **Fetched/imported models now default to disabled (enable on demand)**: bulk imports via "Import from /models" and "Fetch Qoder Models" no longer switch everything on at once — they land in the "Disabled models" area for you to enable as needed, and `/v1/models` now only serves enabled models (a disabled model was previously still listed). Single manual adds stay enabled
+- **CodeBuddy CN listing intl-only models (fixed)**: the CN catalog had picked up the GPT/Gemini family that only the international gateway serves, which failed with "model service info not found" when selected; removed from CN (kept in intl) with a test guard
+- **DeepSeek-V4.1-Flash output ceiling under-reported (fixed)**: its capabilities capped output at 50K (a leftover from the pre-rename id), which truncated client requests; corrected to the model card's 1M context / 384K output / image input / switchable thinking, and the model's two alias ids now carry the same caps
+
+## v1.0.7 (2026-09-07)
+
+### ✨ New
+- **CodeBuddy CN built-in "Total Points" quota row**: the CodeBuddy CN card now leads with a Total Points row summing the live balance of every pack (monthly refills + bonus packs) — no more adding them up yourself; it auto-hides under the "Only with balance" filter once drained; "Bonus Pack" row names are now localized
+- **ZCode usage sync plugin (10router-sync)**: import ZCode's local model-usage ledger into 10Router in one step — providers pointed at 10Router are auto-excluded to prevent double counting; re-runs are idempotent via dedup; imported usage shows up in the usage page and the details tab
+- **Upstream v0.5.69 selective port**: ① Google-family multi-account background refresh → serial + tiered jitter to dodge anti-abuse risk control; ② Anthropic-compatible nodes on real Claude auto-add context-management beta headers (no more silent model-swap); ③ opencode-go stable sessions (free pool no longer trips risk control); ④ Responses parallel tool-call fix (no longer merges N calls into one); ⑤ Claude Fable weekly-quota tracking; ⑥ codex adds gpt-6-astra etc.; ⑦ codebuddy-cn model catalog aligned to server contract
+- **qoder catalog refresh + image passthrough**: model list aligned to server, base64 image direct send; dashboard Antigravity quota grouped by family (per-connection isolation, better than upstream)
+- **Desktop tray update path + leaner menu**: desktop update banner now points to GitHub Releases; tray menu merges check-update / about / start-stop into one compact flow
+- **copilot switched to VS Code extension guide**: no more MITM intercept — three-step extension config (engine layer retained, manually revertible)
+- **Tray icon monochrome**: macOS template icon + Windows dark/light theme black-white adaptive
+
+### 🐛 Fixed
+- **`/responses` root-path auth gap (security)**: prefix added, path now enforces API-key check
+- **Foreign `server_tool_use` poisoning history after Claude combo fallback (400)**: validate `srvtoolu_` prefix and drop foreign tool blocks
+- **MCP deferred tool breaking cache anchors (400)**: anchor now pinned to the last cacheable tool
+- **gemini schema tuple validation 400**: `prefixItems` conversion + array missing `items` placeholder
+- **antigravity system-prompt competitor brand-clean generalization**: OpenCode naming no longer triggers 429
+- **Free-model background-refresh noise + connection-test hang**: de-noised + 15s timeout
+
+## v1.0.6 (2026-09-05)
+
+### ✨ New
+
+- **CodeBuddy CN auto daily check-in** (experimental, off by default): with "CodeBuddy CN auto daily check-in" enabled under Settings → Experimental, the Import / Export buttons on the CodeBuddy CN page are replaced by an automatic daily check-in (each account checks in at a random 00:00–06:00 local time to renew free quota; failures never interrupt the service and a 401 triggers one token refresh + retry) plus a manual "Check in now" button
+- **Antigravity quota aligned with the official site (5h + weekly dual window)**: Antigravity quotas on the usage page now use the official retrieveUserQuotaSummary RPC as the primary source, showing exactly **4 cards** per account (Gemini Models / Claude and GPT models × {5-hour, weekly}), normalized to a 0–100 scale; falls back to the old per-model parse when the RPC is unavailable
+- **OpenCode Free catalog aligned with the official free list**: adds Big Pickle / MiMo-V2.5 / Ling 3.0 Flash Fin / Nemotron 3 Ultra / Nemotron 3.5 Lightning; deliberately excludes the DeepSeek / Laguna free tiers the official docs have dropped (avoids accidental paid calls)
+
+### 🐛 Fixes
+- **Antigravity `ag/gemini-3.8-flash-*` 404**: model addressing now uses per-tier tiered entities and the IDE fingerprint is bumped to 2.11.0 (bare ids 404 with "Requested entity was not found"); adds the un-tiered `ag/gemini-3.8-flash` (routed to the medium tier)
+- **"Only with balance" misses newly-checked-in quota packs**: fixed a case where CodeBuddy CN's daily full packs were hidden after older packs were renumbered and the persistent hide never cleared — the hidden set is now recomputed live so any pack with remaining balance (including fresh full packs) always shows
+- **Friendly message for free-model 429 rate limits**: when a free-tier model (`oc/*-free`, `contributor-free`, APInex `free/`-prefixed, etc.) hits an upstream rate limit and returns 429, the raw English `rate_limit_exceeded` is no longer passed through — instead a friendly Chinese message is returned showing an estimated wait ("try again in ~N sec/min/hr") when the upstream supplied a reset time, otherwise advising to retry later or switch to a paid tier. Paid models and multi-account fallback are untouched
+
+## v1.0.5 (2026-09-03)
+
+### ✨ New
+
+- **Desktop tray app (Windows / macOS)**: an install-and-go desktop app — system tray menu (open console / start / restart / stop / launch at login) plus an embedded window; closing the window minimizes to the tray. Shares data, keys and port with the npm CLI (the two forms run exclusively). Windows ships an NSIS installer (silent `/S` supported) and a portable exe; macOS ships separate Intel + Apple Silicon dmgs
+- **CLI / desktop UI localization**: the npm CLI and the desktop tray now follow the system language (Simplified Chinese / Traditional Chinese / English); override with the `TENROUTER_LANG` env var (`zh-CN` / `zh-TW` / `en`)
+- **Self-serve custom providers + new Skills-page skill**: register a custom OpenAI/Anthropic-compatible endpoint (baseUrl + upstream key + models) at runtime with a dashboard LLM key — no source change, no repackage. The Dashboard **Skills** page gains a **10router-add-provider** card and now links to the `main` branch (previously `master`, which 404'd)
+- **CodeBuddy CN account JSON bulk import / export** (experimental, off by default): after enabling "CodeBuddy CN OAuth import / export" under Settings → Providers, the CodeBuddy CN detail page shows Import / Export buttons to bulk-import or export account credentials in the third-party (wb) JSON format (import dedupes by identity and skips non-CodeBuddy issuers). Import / export now require a dashboard-password confirmation (prevents anonymous token export in no-login mode)
+- **Community welfare providers show by default**: GoRouter / TaBiAI now appear by default (no need to toggle them on), consistent across the provider list / Profile toggle / usage topology
+- **Community welfare providers sort into one group**: in the Free Tier list, community providers (GoRouter / TaBiAI) now cluster into one adjacent block per rank group instead of interleaving with regular freeTier providers by priority/name
+- **New Agnes AI dual-site providers**: international **Agnes AI** (com) + **Agnes AI (CN)** — each with Agnes 2.5 Flash / 2.5 Pro text models (512K / 1M context, vision + reasoning), plus Agnes Image 2.x Flash image models (standard images/generations endpoint; image-to-image / editing)
+- **New "Experimental" settings group**: Profile gains a dedicated Experimental card gathering the developer-oriented default-off toggles (Fetch models from GitHub JSON + CodeBuddy CN import/export) for easy future expansion
+- **New provider APInex (apinex.bond)**: prepaid USD-credit third-party gateway (OpenAI-compatible) with 18 models (13 paid + 5 `free/`-prefixed free models); provider pages show a copyable invite-code chip (APInex: `SLEWP68C`)
+- **Upstream v0.5.65 catalog sync + new search providers**: adds GLM-5.3-Flash(Vision), GLM-4.6V, DeepSeek-V4-Flash-Vision-Exp, Grok-4.6, Claude-Fable-5.1 and more; new search providers Ollama-Search (reuses your local Ollama key) and Xquik (X/Twitter search)
+
+### 🐛 Fixes
+- **Usage table corruption after Cost/Token switch**: toggling the display mode and sorting no longer reverts some rows to currency values; Traditional Chinese usage-table headers are now fully translated
+- **Usage table sorting semantics**: token / cost columns now sort model groups by their totals, matching the header arrow direction
+
+## v1.0.4 (2026-09-01)
+
+### ✨ New
+- **3 new providers**: **TokenBom** (decentralized token marketplace — idle API keys earn credits, credits call many models; 79-model online catalog), **GoRouter** (free gateway, no recharge entry), **TaBiAI** (free gateway, no recharge entry)
+- **Usage history import**: import usage data from 9Router backups (SQLite files) — merged into stats without touching config
+- **Notification overhaul**: global toasts moved to top-center; browser-native alert() dialogs replaced with friendly notifications
+- **Community welfare providers**: GoRouter / TaBiAI are hidden by default with a "community" badge; toggle them on in **Settings → Providers → Show community welfare providers**
+
+### 🐛 Fixes
+- **Provider ordering fixed**: connected providers first, disabled sink to bottom — no longer scrambled by priority
+- **JSON catalog models invisible after enabling**: stale disable records are now cleared ("enabled but not listed" resolved)
+- **B.AI / CodeBuddy CN model catalogs completed**: missing models that caused "not found" when switching are added
+- **Friendly maintenance hints on connection tests**: endpoint down / Cloudflare-blocked shows a maintenance note instead of a misleading "Invalid API key"
+- **Account filter reminder**: quota-pack account filter persists; an amber reminder bar shows when a non-default filter is active
+- **CodeBuddy CN DeepSeek 11150**: DeepSeek-series calls no longer fail with 400 on the reasoning-effort param (auto/off); coding agents (dsh, etc.) work normally
+- **CodeBuddy empty tool name (11133 / unknown tool)**: empty `function.name` in streaming tool calls is normalized, so standard clients no longer mis-detect the tool name
+- **Topology still shows hidden community providers**: the usage topology now honors the "Show community providers" toggle — community welfare sites (GoRouter/TaBiAI) are hidden when it's off
+- **Skills page i18n + Chinese links**: Skills page text is now localized; in Chinese (zh-CN) the links point to the Chinese skill files
+
+## v1.0.3 (2026-08-30)
+
+### ✨ New
+- **4 new providers**: **LongCat** (Meituan), **SenseNova** (SenseTime, free beta), **Dots** (Xiaohongshu Dots Studio, free beta), **B.AI** (aggregator — one key for GPT / Claude / Gemini / DeepSeek / GLM / Kimi / Qwen and more)
+- **Model JSON catalogs for custom providers**: custom nodes can pull an online model list, toggle models on/off individually, and manage them in bulk
+- **fpk update check goes straight to Releases**: fnOS installs jump directly to the matching release download
+
+### 🔒 Security
+- **Progressive login rate-limit**: 5 failed password attempts → 30s / 2m / 10m / 30m lockout
+- **Reject placeholder JWT keys**: copying the public `.env.example` key is ignored; a random key is generated instead
+- **Fix npm package leaking build-machine secrets**: no more keys / machine IDs / data snapshots in the build artifact
+
+### ⚠️ Notes
+- npm package renamed to **`@techysy/10router`** (the old `10router` is an unrelated fork). If you installed `10router-cli`, switch to the new package; data directory unchanged.
+
+## v1.0.2 (2026-08-29)
+
+### 🐛 Fixes
+- **Update check no longer points at a third-party package** (affected 1.0.1): version check, update command, and sidebar install command all target the correct package
+- **postinstall no longer aborts install**: npm install no longer fails from the warm-up script on WSL paths
+
+### ⚙️ Engineering
+- **Test CI added**: tests + regression gate run automatically on push / PR
+
+> ⚠️ **1.0.1 users should upgrade**: its built-in "check for updates" points to an unrelated third-party package.
+
+## v1.0.1 (unreleased; delivered with v1.0.2)
+
+> 1.0.1 was never published as a release (only the npm `10router-cli@1.0.1` briefly existed). The following reached Docker / fpk / standalone users with **v1.0.2**.
+
+### 🔒 Security
+- **4 MITM security fixes**: upstream TLS cert validation restored, root CA private key locked to 0600, no longer blindly killing the process on port 443, auto-cleanup of leftover hosts entries (MITM is off by default)
+
+### ✨ New
+- **Disabled providers sink to the bottom** via a settings toggle
+- **Collapsible desktop sidebar**
+- **OpenCode Go quota usage**
+- **Gitee mirror fallback for model catalogs** (faster in CN)
+
+### 🐛 Fixes
+- **/v1/models no longer returns orphan custom models**
+- **Custom node prefix uniqueness check**
+- **Fix CodeBuddy executor dropping the Agent system prompt**
+
+### ⚙️ Engineering
+- **New npm distribution channel**: `npm i -g 10router-cli`
+- **Generic "fetch models from GitHub JSON" capability** (Fetch Models)
+
+## v1.0.0 (2026-08-26)
+
+### ⚠️ Notes
+1. **Data directory renamed**: `~/.9router/` → `~/.10router/` (Windows: `%APPDATA%\10router`), auto-migrated on first start
+2. **SAML entityID changed**: default issuer is now `urn:10router:sp`; re-register in your IdP
+3. **MITM CA renamed**: re-trust the new CA
+
+### ✨ New
+- **Rebrand**: 9Router → 10Router
+- **i18n multi-region currency**: en / pt-BR / pt-PT / es / de
+- **Multi-platform distribution**: Docker (amd64 / arm64), fnOS fpk (x86 / arm × url / iframe), Standalone