chore(release): v3.19.1

CHANGELOG.md

@@ -5,6 +5,57 @@ All notable changes to CC Switch will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [3.19.1] - 2026-07-31
+
+Development since v3.19.0 is a maintenance pass rather than a feature wave — and the first release in this project's history that deletes more than it adds. Three Chinese Codex gateways move onto native Responses and no longer need local routing takeover: DeepSeek connects straight to `api.deepseek.com` and brings with it a general mechanism for mirroring a vendor's own model catalog verbatim, so the GPT-5 harness and the freeform `apply_patch` registration stay self-consistent instead of collapsing into the neutral template; Volcengine's Ark Coding Plan follows now that the official documentation confirms `/api/coding/v3` serves Responses; and Tencent Hunyuan's TokenHub joins as a new preset. The correctness sweep covers four failures visible in the field: Claude Desktop usage has been counted twice since v3.18.0, once as a proxy row and once as an imported session row (#5938); switching back to the built-in official Codex provider left a third-party key in `auth.json`, so Codex authenticated to the official endpoint with the wrong credential and returned 401 without ever falling through to its login screen; `grok update` failed from Settings with a bare `os error 2` because a GUI-launched app cannot see node; and Grok Build's proxy takeover returned 404 on any non-Responses backend while every request minted a fresh session id (#5677). Rounding it out, nine interface strings that rendered as Simplified Chinese in every language are localized and the Traditional Chinese About page gains the thirty tool-manager strings it was missing, eight models that silently billed at zero gain built-in prices, deep-link import confirmations mask more and truncate less, and 3,166 lines of superseded code plus four npm dependencies are removed.
+
+**Stats**: 12 commits | 71 files changed | +2,324 insertions | -3,680 deletions
+
+### Added
+
+- **Official Vendor Model Catalogs Mirrored for Native Responses Providers**: Codex reads model capabilities from a catalog file, and CC Switch generated every provider's catalog from a neutral template — which is right for an aggregator, but strips capabilities a vendor's own integration depends on. A vendor whose published catalog is bundled with the app is now mirrored verbatim instead. DeepSeek is the first: `src-tauri/src/resources/codex_deepseek_catalog_template.json` carries its `deepseek-v4-flash` and `deepseek-v4-pro` entries with `apply_patch_tool_type: "freeform"`, `web_search_tool_type: "text"`, `supports_search_tool: true`, the low/high/max reasoning levels, and the 17,644-character GPT-5 harness in `base_instructions` and `model_messages` — which has to travel together with the freeform tool registration, because the harness instructs the model to use `apply_patch`. The gate is deliberately narrow: the provider must resolve to the native-Responses profile **and** its `base_url` must be on `deepseek.com`. Matching is by host only and never by model brand, because these entries grant capabilities that an aggregator reselling the same model may not implement. Explicit per-model overrides in the provider's own catalog still win, and an unrecognised model id clones the flagship entry while keeping its own slug. Every other provider profile emits exactly the catalog it did before.
+- **Tencent Hunyuan (TokenHub) Codex Preset**: A "Tencent Hunyuan" entry joins the Codex preset picker under the Opensource Official category, between Bailian and StepFun. It writes `https://tokenhub.tencentmaas.com/v1` with `wire_api = "responses"` and the `disable_response_storage = true` that TokenHub requires, declares `hy3` and `hy3-preview` at a 256K context window (rather than accepting Codex's 128K default), and marks both text-only so `view_image` payloads are never sent to a model that cannot read them. Because it is a native Responses provider, Codex talks to the gateway directly and no local routing is needed; the generated catalog uses the neutral native template, which pins `shell_type = "shell_command"` and drops the freeform `apply_patch` registration that native gateways reject. The address manager and latency test see two candidates from the start — the primary `.com` host and the official `.cn` backup. The region-scoped international site is deliberately excluded, since API keys do not carry across sites. Note that the API key must be a TokenHub key created with the Hy3 scope; Coding Plan and Token Plan subscription keys do not work against this endpoint.
+- **Built-In Pricing for Eight Models That Billed at Zero**: `gpt-5.3-codex-spark`, `gemini-3.5-flash-lite`, `kimi-k2.7-code-highspeed` (2x its `kimi-k2.7-code` base, matching Kimi's Turbo pattern), `glm-5-turbo`, `glm-5v-turbo` and `qwen3.6-flash` had no row in the built-in pricing table and could not be reached by the prefix fallback, so every request against them was recorded at zero cost. Two more rows, bare `claude-opus-4-6` and `claude-sonnet-4-6`, close a subtler gap: model-id resolution strips a date suffix but never adds one, so a log row carrying the undated id matched nothing at all. All eight are seeded insert-if-absent, so a price you already customized is untouched.
+- **Grok Build Joins the Failover Tabs and the Environment-Conflict Check**: The failover settings gain a fourth tab for Grok Build alongside Claude Code, Codex and Gemini, and the startup environment-conflict banner now detects a shell-exported `XAI_API_KEY` or `GROK_DEFAULT_MODEL` — variables that silently override whatever provider is selected. Detection distinguishes exact names from prefixes, so CC Switch's own `GROK_BIN_DIR` and `GROK_HOME` are not reported.
+
+### Changed
+
+- **DeepSeek and Volcengine Ark Coding Plan Connect to Codex Directly, Without Local Routing**: Both presets were declared OpenAI Chat, which made them takeover-required: the provider card carried the "needs routing" badge, switching with the proxy off raised the routing prompt, and every request travelled Codex → local proxy → Responses-to-Chat conversion → upstream. Both vendors now publish official Codex integrations confirming their endpoints serve the Responses API — DeepSeek's `api.deepseek.com` and Volcengine's `/api/coding/v3` — so both presets declare native Responses, the badge and the prompt disappear, and Codex connects straight to the gateway. The generated `config.toml` is unchanged in both cases, since it already emitted `wire_api = "responses"`; what changes is the catalog profile and, for DeepSeek, the context window, which moves from 1,000,000 to the vendor's own 1,048,576. BytePlus deliberately stays on Chat routing until the international site's documentation is verified separately. The Volcengine preset also gained a comment recording a billing rule worth knowing: the pay-as-you-go `/api/v3` endpoint must never be added to this preset's backup addresses, because it bills separately instead of drawing down plan quota.
+- **Catalog `displayName` and `contextWindow` Are Now Explicit-Only**: Both fields previously carried local defaults — the model id and a 128,000-token window — applied before a vendor value could be consulted, so a mirrored catalog would have had its 1M window overwritten with 128K. They are now optional, with the fallbacks applied one level down at entry construction, which makes "left blank" mean "keep whatever the vendor declared". Providers that set the fields explicitly, and every non-mirrored profile, emit identical catalogs to before.
+
+### Fixed
+
+- **Claude Desktop Usage Was Counted Twice**: Claude Desktop traffic through the local gateway landed in the usage dashboard once as a proxy row and once more as an imported transcript row, so its tokens, cost and request counts read roughly double. The regression shipped in v3.18.0: proxy dedup ids were scoped as `session:{app_type}:{provider_id}:{message_id}` for every app except `claude`, which put `claude-desktop` in its own namespace while the transcript importer kept writing the same Claude message under the bare `session:{message_id}` shape with `app_type = 'claude'`. All three dedup defenses failed at once — the primary-key convergence that lets a proxy row absorb an existing session row, the write-side fingerprint probe, and the read-side filter, the latter two both comparing `app_type` with strict equality. The two apps now share the bare namespace again, and the two comparisons are widened by a one-way rule where a `claude` session row can be absorbed by a `claude-desktop` proxy row but never the reverse. Because the read-side filter is also what daily rollups aggregate through, already-stored duplicates stop being counted without any row being rewritten or deleted — see Upgrade notes for the retention limit on that. For Codex, Gemini and OpenCode the widened comparison collapses to the previous exact match, and quota checks keep strict app matching. (#5938, #5951)
+- **Switching Back to the Official Codex Provider Stranded You on a 401 With No Login Screen**: With Codex API-key preservation off (the default), switching to a third-party provider writes that vendor's key into `~/.codex/auth.json`. Switching afterwards to a built-in official provider — whose stored credentials are empty — took the config-only write path, so `config.toml` was replaced while the third-party `OPENAI_API_KEY` stayed on disk. Codex then authenticated to the official endpoint with a foreign key and got 401, and because `auth.json` existed it never fell through to its own login screen, leaving no way out from inside the app. After a successful switch to an official Codex provider, CC Switch now deletes an `auth.json` that contains only an `OPENAI_API_KEY` with no first-class credential beside it — OAuth tokens, a personal access token, an agent identity or a Bedrock key all mark the file as real and leave it untouched, while metadata such as `auth_mode`, `last_refresh` or an account id can no longer shield a stale key. Deleting rather than writing `{}` is deliberate: an empty object resolves to ChatGPT mode without tokens and errors at bootstrap, whereas a missing file yields the login screen. The cleanup runs only after the outgoing provider has been backfilled into the database, so the removed key is preserved and comes back when that provider is selected again. Live-config reads were relaxed in the same change so the post-cleanup state — no `auth.json`, an existing `config.toml` — is no longer reported as "Codex is not installed".
+- **Grok Build Upgrades From Settings Failed With a Bare `os error 2`**: Upgrading Grok Build from Settings → About failed with `Error: No such file or directory (os error 2)` and nothing else. The cause is an asymmetry between how CC Switch probes tools and how it runs them: probing goes through a login shell, which sources the user's rc files and therefore sees nvm, Homebrew and Volta, while lifecycle scripts ran under a non-login shell inheriting the narrow PATH a GUI app is launched with. That normally does not matter, because anchored commands invoke binaries by absolute path — but grok 0.2.112 moved self-update onto npm distribution, so `grok update` now spawns `npm view` and `npm i -g` internally, and npm resolves node through its own shebang. The inner spawn returned ENOENT, which grok surfaced as the bare error. Lifecycle commands on macOS and Linux now run with the login shell's real PATH merged ahead of the inherited one, read through `/usr/bin/env` rather than by echoing the variable, because fish stores PATH as a list and would emit space-separated segments. Native Grok's update additionally chains the official xAI installer as a fallback — deliberately not `npm i -g`, which shares both of the primary's failure modes; the installer is the only node-free path, lands in the same location, and rewrites the CLI's own `installer` setting back to `internal`, healing users whom an earlier npm fallback had moved onto npm distribution.
+- **Grok Build Proxy Takeover Returned 404, and Every Request Looked Like a New Session**: Enabling takeover on a Grok Build provider whose API format was OpenAI Chat or Anthropic produced an immediate 404 with no failover and no usage record — takeover rewrote the base URL and key but left the backend field alone, so the CLI posted to a route the proxy does not register. Takeover now also pins the backend to Responses; the per-provider downgrade to Chat Completions still happens inside the forwarder, and the forced value reverts with the whole live config when the proxy stops. Separately, proxy session extraction recognised only Codex and OpenAI clients, so every Grok Build turn minted a fresh session id marked as not client-provided, which suppressed prompt-cache key injection and per-session grouping in the usage dashboard. Grok's own headers are now read — the conversation id first, then the session id, ignoring the per-request id — under a distinct `grokbuild_` prefix so its rows cannot collide with Codex's. (#5677)
+- **Nine Interface Strings Rendered as Simplified Chinese in Every Language**: Nine strings appeared in Simplified Chinese regardless of the selected language, English and Japanese included. Each call site used the inline-default form with a Chinese literal, but the key existed in none of the four locale files — and i18next resolves a key through the language chain before it considers an inline default, so the English fallback never engaged and the Chinese literal won everywhere. The affected strings cover the Grok Build provider form's validation toast, the failover tooltip shown when an app is not yet taken over, the warning raised when stopping Claude Desktop routing while another app holds takeover and the reason line that explains why routing must start, the duplicate-provider id read failure, the empty Codex common-config error, the routing service's stop and stop-failed toasts, and the "unpriced" cost label the usage tables show for a request that carries tokens but computes to zero. All nine now exist in Chinese, English, Japanese and Traditional Chinese. (#5960)
+- **The Traditional Chinese About Page Fell Back to English in the Tool Manager**: With the interface set to Traditional Chinese, the tool management section of the About page rendered in English — version rows, install and update buttons, result toasts, the install-conflict diagnosis and the entire upgrade-confirmation dialog. The panel was built out across three earlier changes that added labels to Chinese, English and Japanese only, and because i18next falls back to English rather than failing, thirty missing keys were invisible in testing. This gap shipped in every release from v3.16.0 through v3.19.0. All thirty strings are now translated, the install hint was brought in line with the other locales, and a new locale test asserts that every tool-management label exists in all four languages with matching interpolation variables, so this class of drift fails the suite instead of shipping. (#5943)
+- **Seeded Model Prices Disagreed With Vendor List Prices**: Costs are frozen at log time from the built-in pricing table, so a stale seed quietly mis-bills every subsequent request. Four rows were corrected: `deepseek-chat` and `deepseek-reasoner` are now legacy aliases of V4 Flash at $0.14/$0.28 per million input/output with $0.0028 cache read (from $0.27/$1.10 and $0.55/$2.19); `minimax-m3` halves to $0.30/$1.20, the official standard tier; and `gpt-5.6-luna` drops 80% to $0.20/$1.20 while `gpt-5.6-terra` drops 20% to $2/$12, following OpenAI's 2026-07-30 price cut, with `gpt-5.6-sol` deliberately unchanged and the family's cache-write ratio preserved. The repairs rewrite a row only when all four of its cost columns still hold the exact previous built-in values, so a price you edited yourself — or one set by models.dev sync — is never touched.
+
+### Security
+
+- **Deep-Link Import Confirmations Mask More and Truncate Less**: A continuation of the `ccswitch://` confirmation hardening in v3.19.0. Config previews are now built by one shared module that masks secrets recursively through nested TOML tables and JSON objects, which fixes two opposite defects: a Grok Build import rendered no configuration preview at all, while a Codex import printed embedded `api_key` values in the clear. The 300-character truncation on the configuration preview is gone — the full document now renders inside a scrollable box, closing the last place the confirmation could hide part of what it was about to write. Masking itself is stricter everywhere it is used, including the MCP import confirmation: the sensitive-name matcher gained `AUTHORIZATION`, `COOKIE` and `CREDENTIAL` alongside exact matches for `AUTH` and `BEARER`, masked values reveal four leading characters instead of eight, and a value of eight characters or fewer is now replaced entirely rather than shown. Finally, the frontend Base64 decoder no longer trims surrounding whitespace, which could discard a `+` that URL decoding had turned into a space — the same class of frontend/backend decoder divergence fixed in v3.19.0, where the dialog shows one thing and the importer writes another.
+
+### Internal
+
+- **Superseded Code Removed: 3,166 Lines, Fourteen Modules and Four Dependencies**: A deletion pass over code that had no caller. On the Rust side it drops the provider icon inference table, a placeholder health checker, a second never-wired SSE implementation with its own stream and non-stream handlers, two unused proxy session types, and four unreferenced usage parsers along with a dead cost-calculation entry point — the live billing path, its auto-detecting parsers and session id extraction are untouched. Twenty-two `#[allow(dead_code)]` suppressions, which are why the compiler never flagged any of it, go with them. On the front end, fourteen modules with no importer are deleted, including a prompt form modal and a repository manager both superseded by panel rewrites, a duplicate proxy config hook, a circuit-breaker panel that never had an importer in the project's history, and three schema files; their strings are removed identically from all four locales. The Tauri commands behind them are deliberately kept. Four unused npm dependencies are dropped. Two icon-extraction scripts that regenerated the hand-curated icon index are removed, and the index header now states that automatic regeneration is intentionally unsupported. A companion change consolidates proxy status and takeover state onto a single React Query layer — a second, parallel set of hooks over the same commands had zero call sites, so its query keys never had an observer and the invalidations aimed at them were no-ops. Query key strings are byte-identical and the surviving hook keeps its existing poll behavior. (#5916, #5928)
+
+### Upgrade notes
+
+- No database schema migration in this release — `SCHEMA_VERSION` stays at 16, so no pre-migration backup is triggered.
+- Claude Desktop's double-counted usage corrects itself retroactively, but only within the detail-row retention window. The fix suppresses duplicate rows at query time rather than rewriting or deleting anything, so every affected day whose detail rows are still present recovers its correct total on next launch. Detail rows older than 30 days are aggregated into daily rollups and pruned, and a day already rolled up by a build without this fix keeps its inflated figure permanently. The regression entered with v3.18.0 (2026-07-21).
+- The eight newly priced models are re-costed retroactively where possible: startup backfills requests that were logged with zero cost, so dashboard totals for those models will rise. Rows already rolled up and pruned stay at zero. The four repriced models work the other way — backfill only touches zero-cost rows, so requests already logged keep the old rate and only new requests bill at the new one; historical and current spend for the same model will differ. Manual price edits, models.dev sync values and deletion tombstones are re-applied after seeding and always win.
+- Preset changes apply only to newly created providers. A DeepSeek or Volcengine Ark Coding Plan provider already saved keeps its stored API format, keeps requiring local routing, and keeps its old catalog; re-create it from the preset, or switch the API format in the provider form, to get the direct connection. A provider that is already on native Responses with a `deepseek.com` base URL picks up the mirrored official catalog on its next switch without any re-save, because the gate reads the live configuration.
+- `deepseek-v4-pro` cannot be used over the direct connection yet. It stays listed in the preset, and in the mirrored vendor catalog, but DeepSeek has not opened its Codex integration for that model — early August 2026 by their own estimate — so selecting it while the provider is on native Responses fails upstream. Use `deepseek-v4-flash`, which is the preset default and unaffected. To use pro today, switch that provider's API format back to OpenAI Chat and enable local routing takeover: that is the path DeepSeek took before this release, and the proxy's Responses-to-Chat conversion serves pro exactly as it did.
+- The mirrored DeepSeek catalog declares a minimum Codex client version of 0.144.0, which CC Switch does not verify — the freeform `apply_patch` registration it carries needs that release or newer. The generated catalog file also grows to roughly 75 KB for the two mirrored models, since each carries the full harness text.
+- Because DeepSeek, Volcengine Ark Coding Plan and Tencent Hunyuan no longer require takeover, their traffic can bypass the local proxy entirely, and per-request proxy usage accounting does not see it. The usage itself is neither lost nor indistinguishable — Codex session-log import still records it — but that path carries no provider identity: all Codex usage that did not go through the local proxy is grouped under a single `Codex (Session)` entry, official-subscription consumption included. So a provider moving from routed to direct also moves its usage out from under its own name. Model attribution is unaffected: every row keeps its model id, and the usage panel's per-model table still separates `deepseek-v4-flash`, `hy3`, `ark-code-latest` and the official GPT models into their own rows with their own token and cost figures. Keep local routing takeover enabled only if you need the provider dimension specifically — for instance to compare the same model across several aggregators.
+- The stale Codex `auth.json` cleanup is gated on the incoming provider carrying the explicit official category and on the outgoing provider being backfilled successfully. An official entry created by hand without that category, or a switch whose backfill failed, still leaves the residue behind.
+- Enabling proxy takeover on a Grok Build provider now rewrites its backend to Responses in the live configuration. The stored provider row is untouched and the live file is restored from its backup when the proxy stops.
+- On macOS and Linux, every tool install and update triggered from Settings → About now runs with the login shell's PATH merged ahead of the inherited one, so a binary a lifecycle script resolves by name may resolve differently than before; each action also spawns one extra shell to read that PATH, which executes the user's interactive startup files. Windows is unchanged. Users on grok 0.2.112 or later may see two enumerated installations — the native one plus a global npm package that `grok update` created itself; they are kept in sync upstream and report the same version.
+- Environment-conflict detection for Claude Code, Codex and Gemini moved from substring to prefix matching, so variables that merely contain an app's name — `MY_ANTHROPIC_API_KEY`, `OLD_GEMINI_API_KEY` — are no longer reported as conflicts.
+- Deep-link import confirmations now reveal less of each secret than before: four leading characters instead of eight, and values of eight characters or fewer are masked entirely.
+
 ## [3.19.0] - 2026-07-30
 
 Development since v3.18.0 is headlined by a security-hardening wave and a proxy correctness fix with immediate field impact. On the security side (#5811 and follow-ups): skill installs from GitHub repositories are hardened against zip-slip and repository-coordinate path traversal with hard archive limits; a Gemini common-config credential leak is closed and a one-time startup cleanup scrubs keys that already leaked into other providers' configs; imported SQL backups now run under a SQLite authorizer that denies `ATTACH` and every other statement that can reach outside the import database; common-config snippet handling no longer follows `__proto__` into the global prototype; terminal launch escapes project paths with POSIX single quotes so a directory name can no longer inject commands; and `ccswitch://` import confirmations show every argument, env var, URL, and the complete usage-script body — nothing hidden by truncation, nor by a Base64 variant the dialog could not decode, credential-shaped values masked — flag risky values, and import usage scripts disabled by default. On the proxy side, tool-result images are lifted out of stringified tool text and re-emitted as native media on every conversion bridge, ending the ~9,000× token inflation that pushed Codex sessions past their context window after a few screenshots (#4465, #5663). Usage statistics gain opt-in automatic price sync from models.dev with a persistent local override file (#5734), plus coverage of Grok CLI's official OAuth mode imported from session logs and SuperGrok subscription quota on provider cards. Rounding it out: in-app updates are served from a Cloudflare R2 mirror at `dl.ccswitch.io` with GitHub as fallback, Codex usage import reuses parsed parent rollouts across forked sessions (#5626), preset defaults advance to Claude Opus 5, GPT-5.6 Sol, and Gemini 3.6 Flash, the OpenClaw Kimi For Coding base URL is corrected, and the GPT-in-Claude-Code routing guide is now available in English and Japanese.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "cc-switch",
-  "version": "3.19.0",
+  "version": "3.19.1",
   "description": "All-in-One Assistant for Claude Code, Codex & Gemini CLI",
   "type": "module",
   "scripts": {

src-tauri/Cargo.lock

@@ -758,7 +758,7 @@ dependencies = [
 
 [[package]]
 name = "cc-switch"
-version = "3.19.0"
+version = "3.19.1"
 dependencies = [
  "anyhow",
  "arboard",

src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "cc-switch"
-version = "3.19.0"
+version = "3.19.1"
 description = "All-in-One Assistant for Claude Code, Codex & Gemini CLI"
 authors = ["Jason Young"]
 license = "MIT"

src-tauri/tauri.conf.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "CC Switch",
-  "version": "3.19.0",
+  "version": "3.19.1",
   "identifier": "com.ccswitch.desktop",
   "build": {
     "frontendDist": "../dist",