ClewdR: baixar versions anteriors

Release Notes
Security
- Distrust compromised emSign certificate chains. Update the bundled CA
certificates (webpki-root-certs 1.0.9 → 1.0.10), removing
emSign ECC Root CA - G3. This responds to the disclosure of intermediate
CA private keys belonging to India's eMudhra/emSign and prevents ClewdR's
bundled trust store from accepting chains through those intermediates.
The update protects upstream HTTPS connections against public-network
man-in-the-middle tools using the disclosed keys to forge certificates.
- Upgrade recommended. Install v0.13.6 and restart ClewdR; container users
should pull the updated image and recreate their containers. Updating the
operating system's CA bundle alone does not update the roots embedded in
existing ClewdR binaries.
References: RUSTSEC-2026-0383 / CVE-2026-111943,
upstream root removal.
安全更新
- 因印度 CA 机构 eMudhra/emSign 的中间 CA 私钥泄露，将内置 CA 包
webpki-root-certs 从 1.0.9 更新至 1.0.10，移除
emSign ECC Root CA - G3 信任根，防范公网中间人工具利用泄露密钥
伪造证书、拦截 ClewdR 的上游 HTTPS 通信。
- 建议升级到 v0.13.6 并重启；容器用户请拉取新镜像并重建容器。
仅更新操作系统 CA 包不会更新已有 ClewdR 二进制内嵌的信任根。

Release Notes
Fixes
- Claude Code 2.1.258 compatibility. Requests now identify as the current
client instead of the legacy 2.1.76 release. Billing attribution also joins
JavaScript UTF-16 code units 4, 7, and 20 before UTF-8 hashing, keeping the
cc_version suffix correct for emoji and surrogate-pair boundaries.
- Portable GNU/Linux archives. Nix-built GNU binaries have their loader
path repaired before packaging, so the release ZIPs run outside the Nix
store on supported distributions instead of failing with “No such file or
directory”.
- Container publication. Multi-architecture image publishing now preserves
the existing sha-<7> tags, produces real amd64/arm64 manifest lists, and
avoids publishing images from manual workflow dispatches.
Changes
- Reproducible release builds. Linux GNU, musl, Android, frontend and
distroless container outputs now use pinned Nix definitions. nix develop
provides the matching stable/wasm/nightly/trunk toolchain, while Windows and
macOS retain their native build paths. The HTTP stack and other dependencies
were refreshed, including the stable wreq 0.16 release.
- Faster, verifiable CI. Linux targets and both container architectures now
build in parallel; exact feature sets, shared frontend store paths and
minimal per-target toolchains avoid redundant compilation. Workflow YAML and
the image-push/tag scheme are also checked locally through xtask before
publication.

Release Notes
Fixes
- Quoting a value in the environment locked you out. CLEWDR_ADMIN_PASSWORD="12345" set the password to "12345" with the quotes attached, so the password printed at startup was not the one the login page would take. Compose files and .env files pass quotes through as data rather than stripping them the way a shell does, so writing them is a common habit, and every earlier version quietly absorbed it. A value written as one quoted string is unwrapped again, and surrounding whitespace is dropped again. CLEWDR_PASSWORD was affected the same way. (#157)
A value that only contains quotes — say "hi", or two of them side by side — still keeps every character, and single quotes are still part of the value. If a password genuinely needs to begin or end with a space, wrap it in double quotes.
Setting a password that reads as a number, CLEWDR_PASSWORD=12345, still works. That was fixed in 0.13.3 and is unaffected; the quoting behaviour was removed alongside it by mistake, and only the quoting is being restored.

Release Notes
Fixes
- Beta features a client asked for were dropped on the way through. ClewdR sent its own anthropic-beta header and discarded whatever the client had set, so anything gated behind a beta flag the client requested — context management among them — never reached Anthropic. The client's flags are now merged with ClewdR's own and passed on, without duplicates.
- Streamed responses failed to parse when citations were enabled. The list of delta types a response could contain was missing citations_delta and compaction_delta, both of which the API sends today, and had no fallback for anything unrecognised. The failure landed mid-response, after the reply had started arriving, where nothing already sent could be taken back. Both are now understood, and a delta type ClewdR does not know is passed through untouched rather than ending the response.
- Six request fields were accepted and then silently dropped: cache_control, inference_geo, speed, diagnostics, fallbacks and fallback_credit_token. A client setting any of them got a reply that ignored it, with nothing to indicate why. They are forwarded now, checked field by field against Anthropic's own SDK.
- The cookie listing sent OAuth access and refresh tokens to the browser. The tokens were serialized along with the rest of each cookie's status, and the page discarded them on arrival — so they crossed the wire, sat in browser buffers and any proxy log on the path, and were never read. The endpoint requires the admin password, so this was not a way in for anyone unauthenticated; what it changed is how much an intercepted admin session gives up. They are removed before the response is built.
Performance
- Every request spent about 60ms building a tokenizer it then used for 0.04ms of work. Token counting rebuilt the 200,000-entry encoding table from scratch on each call rather than keeping it, and both request and response counting did this, so a request that counted input and output tokens paid it twice. The table is now built once per process. Counts are unchanged.
Development
- The Anthropic request and response types live in their own crate, anthropic-wire, instead of sitting inside the server. Nothing about the protocol depends on ClewdR, and separating them means the compiler enforces that. It also surfaced a conversion broad enough to turn any string into an assistant message, which had gone unnoticed while everything shared one crate.
- Eight direct dependencies are gone: oauth2, figment, moka, serde_with, itertools, trie-rs, passwords and const_format. Each was doing something small enough to write directly — two OAuth grants, a config merge, two caches, three serde adapters, one message fold, a stop-sequence match, a password generator, one string concatenation. The lockfile is down from 453 packages to 389.
- Release builds no longer compile the frontend for the host on top of the target build, and CI installs its tools with --locked. A prebuilt binary download that timed out was falling back to building from source with freshly resolved dependencies, which picked up an incompatible pair and failed the frontend job on a commit that had not touched the frontend.

Release Notes
Fixes
- The cookie status page failed to load, showing invalid type: integer ... expected a string and a total of 0 cookies. Reset boundaries are held as timestamps internally, and were sent that way whenever the live usage lookup did not answer, where the interface expects the written-out form. The page reads the whole listing as one piece, so a single cookie the lookup could not reach hid every other cookie with it. Boundaries that are already known are now shown even when the lookup fails, instead of the field going blank.
- A non-ASCII admin password left the web interface blank. The saved token is shortened for display by keeping a few characters at each end, but the length it was checked against was measured in bytes, so a password of four Chinese characters looked long enough to shorten and the cut landed inside a character. The page went down as it loaded, and because the token is read back from browser storage, reloading did not help. Shortening now counts characters throughout.
- The configuration file could be truncated when two saves overlapped. It was rewritten in place, so anything reading it — including ClewdR's own startup — could catch it after it had been emptied and before it was filled again. Saves now write a temporary file and rename it into place, which cannot be observed half-finished, and the file is created with owner-only permissions rather than having them applied after the contents are already on disk.
- A cookie whose cooldown had expired was returned to rotation without being recorded, so restarting in that window listed it as exhausted until the next request moved it back.
Docker
- The image builds against wreq 6.0.0-rc.29 again. 0.13.1 had to hold it at rc.28, because rc.29 changed its TLS backend to one that carries BoringSSL as C++ and the image had no C++ compiler for musl. What was wanted then was a musl toolchain that CI could reach; it turned out to already exist as a published image, so the build now starts from one that carries a complete musl GCC.
- Image builds reuse cached layers between runs. Caching had been switched off, so every build recompiled all dependencies, BoringSSL included.
- Pull requests build the image for both architectures but no longer publish it. Builds from forks failed at the registry login, which cannot succeed for a fork, and every other pull request was pushing an image nobody asked for.
Development
- Replaced the cookie actor with a mutex-guarded pool, removing the ractor dependency. None of the actor's handlers ever awaited, so the framework was supplying serialized access to shared state and nothing more.
- The pool now owns the cookies outright. They used to be copied back into the global configuration on every change, which is why saving settings had to carry them across by hand to avoid wiping them.
- Dropped struct_iterable, rustls and aws-lc-rs, none of which were used. The latter two were left behind by the removal of Gemini support; HTTP goes through wreq, which brings its own TLS.
- Pull requests no longer cross-compile the full release matrix. Nine binaries were built, uploaded and discarded a day later on every pull request. A single host build now covers the release profile and the embedded frontend, and the matrix runs on tags, where the artifacts are actually used.

Release Notes
Fixes in 0.13.1
- Fixed the Docker image, which failed to build for 0.13.0 and was therefore never published for that version. Everything below shipped in 0.13.0 and reaches Docker users for the first time here; the 0.13.0 binaries were unaffected.
Models
- Added Claude Fable 5, Opus 5, Opus 4.8, Opus 4.7, Sonnet 5 and Haiku 4.5 to the model list.
- Removed models Anthropic has retired: Claude 3.7 Sonnet, Sonnet 4 and Opus 4. Requests to these fail upstream, so listing them only misled clients.
- Removed Opus 4.1, which Anthropic retires on 2026-08-05.
- Every model is still offered in a -thinking variant.
Request handling
ClewdR now adjusts requests to match what the target model accepts, instead of forwarding parameters that are rejected with a 400.
- temperature, top_p and top_k are dropped on Claude 4.7 and later, which no longer accept them, and on any model where thinking is active.
- temperature and top_p are no longer sent together, which Claude 4.1 and later reject.
- Thinking is converted rather than dropped: a token budget becomes adaptive thinking on models that only support adaptive, and adaptive becomes a budget on models that only support the older form.
- output_config.effort is stepped down to the nearest level the model implements, and xhigh is only sent to Claude 4.7 and later.
- Thinking budgets are held within the range the API accepts, rather than being passed through as given.
Fixes
- OpenAI-compatible requests using reasoning_effort failed with a 400. The value was being sent as a thinking budget, and low mapped to 256 tokens, below the API minimum of 1024. Effort levels are now mapped properly, including xhigh and max.
- The -thinking suffix now also requests the reasoning text on Claude 5 models, which otherwise think without returning it.
Frontend
- Increased the interface font size. The root size was below the browser default and the smallest labels rendered at about 10px, or under 10px on mobile. Nothing is now smaller than 12px, and the interface follows the browser's own font size setting.
Also shipped in 0.12.26
These were released in 0.12.26, but its release notes did not mention them:
- The web interface was rewritten in Leptos and WebAssembly, replacing the React frontend (#149).
- thinking.display is preserved when forwarding requests upstream (#151).
- Improved light theme contrast.
- Removed the 7-day Opus quota tracking.
Development
- Added cargo xtask for building, running and linting: cargo xtask dev runs the backend and the frontend with hot reload, and cargo xtask ci runs everything CI runs.
- Enabled clippy::pedantic across the workspace, and CI now checks formatting and lints.
- Rewrote both READMEs. Several documented details, including the download command and parts of the web interface description, did not match the software.
Compatibility
- Clients pinned to Claude 3.7 Sonnet, Sonnet 4, Opus 4 or Opus 4.1 must move to a current model. /v1/models lists what is accepted.
- Sampling parameters sent to Claude 4.7 and later are now dropped rather than forwarded. Previously these requests failed; they now succeed without those parameters.
- Building from source now requires the frontend to be built first. Use cargo xtask build.

Release Notes
Models
- Added Claude Fable 5, Opus 5, Opus 4.8, Opus 4.7, Sonnet 5 and Haiku 4.5 to the model list.
- Removed models Anthropic has retired: Claude 3.7 Sonnet, Sonnet 4 and Opus 4. Requests to these fail upstream, so listing them only misled clients.
- Removed Opus 4.1, which Anthropic retires on 2026-08-05.
- Every model is still offered in a -thinking variant.
Request handling
ClewdR now adjusts requests to match what the target model accepts, instead of forwarding parameters that are rejected with a 400.
- temperature, top_p and top_k are dropped on Claude 4.7 and later, which no longer accept them, and on any model where thinking is active.
- temperature and top_p are no longer sent together, which Claude 4.1 and later reject.
- Thinking is converted rather than dropped: a token budget becomes adaptive thinking on models that only support adaptive, and adaptive becomes a budget on models that only support the older form.
- output_config.effort is stepped down to the nearest level the model implements, and xhigh is only sent to Claude 4.7 and later.
- Thinking budgets are held within the range the API accepts, rather than being passed through as given.
Fixes
- OpenAI-compatible requests using reasoning_effort failed with a 400. The value was being sent as a thinking budget, and low mapped to 256 tokens, below the API minimum of 1024. Effort levels are now mapped properly, including xhigh and max.
- The -thinking suffix now also requests the reasoning text on Claude 5 models, which otherwise think without returning it.
Frontend
- Increased the interface font size. The root size was below the browser default and the smallest labels rendered at about 10px, or under 10px on mobile. Nothing is now smaller than 12px, and the interface follows the browser's own font size setting.
Also shipped in 0.12.26
These were released in 0.12.26, but its release notes did not mention them:
- The web interface was rewritten in Leptos and WebAssembly, replacing the React frontend (#149).
- thinking.display is preserved when forwarding requests upstream (#151).
- Improved light theme contrast.
- Removed the 7-day Opus quota tracking.
Development
- Added cargo xtask for building, running and linting: cargo xtask dev runs the backend and the frontend with hot reload, and cargo xtask ci runs everything CI runs.
- Enabled clippy::pedantic across the workspace, and CI now checks formatting and lints.
- Rewrote both READMEs. Several documented details, including the download command and parts of the web interface description, did not match the software.
Compatibility
- Clients pinned to Claude 3.7 Sonnet, Sonnet 4, Opus 4 or Opus 4.1 must move to a current model. /v1/models lists what is accepted.
- Sampling parameters sent to Claude 4.7 and later are now dropped rather than forwarded. Previously these requests failed; they now succeed without those parameters.
- Building from source now requires the frontend to be built first. Use cargo xtask build.

Release Notes
Changes
- Remove the -1M and -1M-thinking model aliases from the model list and request handling.
- Stop forwarding client-provided anthropic-beta values for long-context requests. The OAuth beta header required by Claude Code authentication remains enabled.
- Remove the special long-context 429 handling; these responses now use the standard rate-limit path.
- Remove the obsolete 1M toggle styles from the frontend.
Compatibility
- Clients should use the standard Claude model IDs without the -1M suffix.
- Client-provided long-context beta headers are no longer passed to Anthropic.

Release Notes
What's New
- Use native incognito mode (is_temporary) for non-preserved chats (PR #145 by @GottenHeave)
- Web usage endpoint for enterprise accounts
- Update TLS emulation to Chrome 145
Bug Fixes
- Fix clewdr.toml file permissions on Unix: now created with 0600 instead of default umask (#122)
- Add fallback for unknown ContentBlock types to prevent 422 deserialization errors (#97)
- Fix OAI ImageUrl to Claude Image format conversion in Claude Code proxy (PR #121 by @DragonFSKY)
- Fix enterprise usage tracking (PR #144 by @GottenHeave)
- Work around a bug in tower-serve-static (#147)
Improvements
- Unify HTTP client construction across codebase
- Always use mimalloc as default allocator
- Unpin tracing-subscriber, allow ANSI color output
- Update dependencies to latest versions
- Use distroless Docker image

Release Notes
- Fix: Switch default allocator to mimalloc to address memory issues with long-running processes.
- Update dependencies to latest versions for improved performance and security.

Release Notes
- Fix: added the Adaptive variant to the Thinking enum.
- Exposed custom_system in the web Settings UI, and clarified that custom_prompt is only for the Claude Web prompt field.
- Claude Code now emulates claude-code/2.1.76 more closely, including User-Agent: claude-code/2.1.76 on Claude Code requests.
- Claude Code requests now prepend a 2.1.76-compatible billing system header before custom_system, using sampled characters from the first user message and fixed cch=00000.

Release Notes
- Requests now keep the original system content from clients unchanged by default (except existing cache-control scope cleanup).
- Kept custom_system support: when configured, it is prepended to the request system content.
- Dropped the system field when it is empty (null, blank string, or empty system block list) before forwarding requests.

Release Notes
2026-02-28
- Removed automatic prelude injection into system prompts for non-Claude Code requests.
- Requests now keep the original system content from clients unchanged (except existing cache-control scope cleanup).

Release Notes
- fix: remove top_p when temperature is set

Release Notes
- build(docker): stop using musl targets in Docker image build; compile native GNU target directly to avoid boring-sys2 open64/fopen64/stat64 link failures
- build(docker): switch runtime base image to debian:trixie-slim and include required runtime libs (ca-certificates, libgcc-s1, libstdc++6)
- ci(docker): disable buildx cache for Docker publish workflow (no-cache: true) to avoid stale-layer interference during multi-arch builds

Release Notes
- chore(deps): bump wreq to 6.0.0-rc.28 and wreq-util to 3.0.0-rc.10
- fix(code-auth): restore Claude Code OAuth authorize reliability by explicitly sending cookie + claude-code user-agent in code exchange
- fix(code-auth): when upstream returns permission_error: Invalid authorization, clear cached token and retry OAuth flow automatically
- fix(compat): adapt request/client usage for wreq v6 API changes

Release Notes
- fix: stop stripping context-1m beta token at the 200000 max_tokens threshold
- feat: add Sonnet/Opus 1M tri-state mode (auto-probe / enabled / disabled) with probe-fail auto fallback and auto-disable
- feat: restore OpenAI -1M model aliases for Sonnet 4/4.5/4.6 and Opus 4.6
- fix: limit 1M probing to Sonnet 4.x and Opus 4.6 lanes
- fix: split 429 handling for 1M flow (long-context gate 429 falls back without cooling cookie; normal rate-limit 429 still cools cookie)

Release Notes
- feat: strip context-1m beta token when max_tokens <= 200000

Release Notes
- feat: forward inbound anthropic-beta to Claude Code upstream while always appending oauth beta
- refactor: remove -1M model suffix handling and always use the base Anthropic beta header
- refactor: remove Claude 1M auto-probing and related config/cookie/frontend fields

Release Notes
- fix: update ClewdrCookie regex patterns for enhanced validation and add tests for longer cookies

Release Notes
- feature: split 1M probing state into Sonnet and Opus lanes (cookie-level tracking)
- feature: add configurable auto-probe toggles for Sonnet 1M and Opus 4.6 1M
- ui: show separate Sonnet/Opus 1M status badges in Cookie details
- ui: arrange API config toggles into 3 rows x 2 columns to reduce crowding

Release Notes
- fix: increase default body size limit to 32MB
- feature: add claude-opus-4-6 model support
- feature: add experimental 1M probing support for claude-opus-4-6

Release Notes
- ci: restore x64-musllinux in ci (use cross tools)
- feature: restore santizer and add an option for it

Release Notes
- feat: replace RequireXApiKeyAuth with RequireFlexibleAuth fully to support bear token
- chore: increase cookie log ellipsis length to 20 chars

Release Notes
- remove santizer
- remove musllinux temporalily to be built
- update to adapt cookie like sk-ant-sidxx-

Release Notes
- Fix toolcall param bug
- Expand Claude message schema support (tool_choice strings, content block variants, image sources, stop reasons)

Release Notes
- Fix a custom tool bug

Release Notes
- Improved Claude Code request handling with User-Agent detection (including claude-cli) and a configurable prelude system prompt for non-Claude Code clients.
- Fixed built-in tool schema handling (including web_search_20250305) to make web search/tooling work reliably.
- Fixed ANSI color output: disable escape codes when stdout is redirected and in log files; adjusted tracing-subscriber version for correct terminal ANSI behavior.
- Renamed “NonPro” to “Free” and updated related UI/config messages (e.g. “Skip Free”).
- Updated dependencies and removed unused helper files/scripts; stopped ignoring .claude via .gitignore.

Release Notes
Unreleased
- Removed database persistence support and related storage APIs; configuration now always saves to clewdr.toml.
- Dropped Gemini AI Studio and Vertex AI routes, providers, and credentials management from the backend.
- Simplified the frontend by removing Gemini tabs, key/vertex upload flows, and database storage controls.
- Updated documentation to reflect the Claude-only focus going forward.

Release Notes
Maintenance
- Removed database persistence support and related storage APIs; configuration now always saves to file.
- Dropped Gemini AI Studio and Vertex AI routes, providers, and credentials management from the backend.
- Simplified the frontend by removing Gemini tabs, key/vertex upload flows, and database storage controls.
- Updated documentation to reflect the Claude-only focus going forward.

Release Notes
Feature
- 通过 CLAUDE_CONSOLE_ENDPOINT 环境变量 可以设置 CC_REDIRECT_URI 的值 实现 oauth 的反代设置

Release Notes
Fix bug
- Fix Claude Code usage counters not resetting to zero when session/7-day cooldown windows expire without a new request (session, weekly, Opus, Sonnet buckets now clear on periodic checks).

Release Notes
Maintenance
- Fix opus 4.5 parameter problem

Release Notes
Maintenance
- Add 4.5 Opus thinking support for openai compactible format

Release Notes
Bug Fixes
- Fix auto-update GitHub release lookup by stripping email from author metadata so the GitHub API URL points to the right repo
- Correct Android-only function call in build.rs so Android builds compile again
Maintenance
- Update Cargo and frontend npm/pnpm dependencies to latest compatible versions

Release Notes
Bug Fixes
- Update Claude Code OAuth token refresh to use api.anthropic.com with API-style headers for better compatibility

v0.11.21
Bug Fixes
- Fix double slash problem for Claude Web API endpoints

v0.11.21
Bug Fixes
- Fix double slash problem for Claude Web API endpoints

Bug Fixes
- Fix double slash problem for Claude Web API endpoints

Highlights
- Claude tool-result content now accepts structured array payloads, keeping MCP clients such as OpenHands compatible with the proxy.
Improvements
- Normalized endpoint constants and switched to Url::join when building Claude Code, Claude Web, Gemini, and OAuth URLs, avoiding double-slash mistakes when overriding API hosts.
- Refined Claude message sanitization so blank assistant turns are dropped without discarding legitimate content blocks.
Bug Fixes
- Fixed Claude Code chat and token-count requests failing when the configured endpoint already included a trailing slash.
Dependency Updates
- Bumped zip to 6.x and enable-ansi-support to 0.3, refreshing the lockfile to the latest compatible releases.

v0.11.18
Bug Fixes
- Claude API: trim whitespace-only assistant messages before dispatching so Anthropic no longer returns 400 for streams that include blank turns.

Bug Fixes
- Claude API: trim whitespace-only assistant messages before dispatching so Anthropic no longer returns 400 for streams that include blank turns.

Bug Fixes
- FIX A 1M AUTO DETECT BUG FOR PRO SUBSRIPTION

Features
- Claude 4.5 support
- Update model prefixes and request handling for Claude 3.7 Sonnet (2025-02-19).
- Add token usage tracking and count_tokens endpoints (Code/Web).
- Add config toggle to enable web count_tokens.
- Cookie management enhancements: 1M-context capability flag and visualization; current/total token usage fields.
- Cookie quota visualization
- /api/cookies now returns ephemeral quota utilization: 5-hour session, 7-day, and 7-day Opus (percent).
- Also returns corresponding reset timestamps (ISO 8601). Frontend shows expandable details.
- Gemini support and management
- AI Studio and Vertex proxy, providing OpenAI-compatible chat/completions.
- Admin APIs: submit/delete Gemini keys; list/add/delete Vertex service account credentials.
- Frontend Gemini tab: key submission/visualization; Vertex credential list and upload.
- Persistence and admin
- Add storage Import/Export/Status endpoints (available when built with db feature).
- Settings page shows storage status summary and provides Import/Export (enabled in non-file modes).
Improvements
- Ensure log directory exists before writing when file logging is enabled.
- Install aws-lc crypto provider before using rustls to improve TLS stability.
- Settings UI: remove legacy Vertex Settings section; update related copy and translations.
- Refine DB initialization and storage health reporting; tidy provider abstractions and router structure.
Bug Fixes
- Docs: clarify private key handling for Google Cloud service account JSON.

Features
- Add support for database persistence using SeaORM (optional feature).
Bug Fixes
- Correctly handle 403 errors for CC
Improvements
- Improve build process

Bug Fix
- Create app dir in XDG directory if it doesn't exist

Features
- Support Claude 4 Sonnet 1M context in Claude Code.
- Support XDG Base Directory Specification

Bug Fixes
- Fix thinking field default fallback handling for Claude

Build
- Add WoA target in GitHub Action
- Optimize for binary size in Cargo.toml
- Enable UPX compression in Dockerfile and GitHub Actions

- Fix Docker build workflow to properly handle version tags

Features
- feat: add optional support for mimalloc as a global allocator
Bug Fixes
- fix: handle 401 status code by returning Null reason in CheckClaudeErr implementation
- fix: refactor prelude text handling in system messages for Claude Code
- fix: remove tag trigger from push event in Docker build workflow

Fix
- Bug: Fix issue with response headers not being checked correctly in overloaded responses.

Fix
- Dep: Remove mimalloc for minimal memory footprint at the cost of performance.

Fix
- Gemini: Clear unsupported frequency_penalty field from requests.
- Claude: Handle overloaded responses more gracefully.
Improvements
- Build: Set up optional feature flags for faster build, smaller size and lower memory usage.
- Action: Refactor GitHub Actions workflows for better maintainability.

Fix
- GitHub Actions: Update version and latest tags in Docker build workflow to ensure proper tagging of Docker images.

Fix
- Update Check: Fix rolling log creation in with no_fs option.
Feature
- Docker Build: Add support for multi-arch builds (amd64, arm64).

Fix
- Update Check: Skip update check if no_fs feature is enabled to avoid unnecessary network requests and potential errors in environments where filesystem access is not available.
v0.11.0
Features
- OpenAI API Transform: Add transforms_json function to convert Claude API responses to OpenAI format with proper usage mapping and finish reason conversion
- No-FS Configuration: Replace compile-time no_fs feature with runtime configuration option for flexible deployment scenarios
- Enhanced Logging: Automatic log directory creation and improved file logging setup with tokio console support
- Thinking Mode Support: Add proper thinking mode configuration with budget token support and type validation
Improvements
- Request Processing: Simplify Claude Web request transformation by removing API format branching and unused OpenAI compatibility layer
- UI Cleanup: Remove padding text configuration from frontend components and related localization entries
- State Management: Refactor middleware state handling with improved generic type support and cleaner FromRequest implementations
- Response Handling: Unify response processing using forward_response utility and remove duplicate response transformation logic
- Performance: Replace chrono with stdtimeInstant for more accurate request timing measurements
- Code Organization: Restructure Gemini middleware path handling and improve import organization
Fixes
- Dependencies: Update multiple dependencies including clap, h2, ractor, security-framework, and zerovec to latest compatible versions
- Docker: Complete overhaul of Dockerfile using distroless images, musl targeting, and optimized alpine-based cargo-chef builds
- Licensing: Add proper AGPL-3.0 license declaration in Cargo.toml
- API Compatibility: Fix return types across Claude API endpoints and correct error response type conversions
- Frontend Dependencies: Update frontend package dependencies and resolve pnpm lock file conflicts
Removed
- Debug Dockerfile: Remove separate debug Dockerfile in favor of unified build process
- Padding System: Remove token padding configuration and random text generation functionality
- API Format Handling: Remove dual API format support in favor of unified Claude API handling

Feature Updates
- Update Claude Opus 4.1
Removed Features
- Cache responses, as few users use it and hard to maintain.

Fix
- Special handle for Claude Sonnet 4 model
TODO
- Major refactor ongoing. Will be released in v0.11.0.

Features
- Usage Tracking: Added comprehensive usage tracking for Claude API requests and responses with token counting
- Enhanced Response Handling: Improved response handling by adding usage information and token counting capabilities
Fixes
- Configuration Management: Improved config saving with async handling and enhanced error logging
- Cache System: Simplified cache system by removing TTL handling for Claude Code prompts and updated to use time-to-idle instead of time-to-live
- Dependencies: Updated async-compression to version 0.4.27 and crc32fast to version 1.5.0, removed unused scopeguard and figlet-rs dependencies
- Error Handling: Enhanced error handling for token requests and improved cookie request handling
- Web Admin: Corrected web admin endpoint URL formatting and adjusted import order
- Serialization: Added default serialization for credential field in VertexConfig and reordered imports in GeminiState
Documentation
- README: Improved formatting and clarity in README and README_zh files
- Icons: Added navigation SVG icon to README and updated SVG icon styling
- License: Removed incorrect MIT license reference from README
- Serialization Docs: Updated README for serialization and import changes
Internal Improvements
- Simplified banner generation by removing figlet-rs dependency
- Improved asynchronous handling throughout the codebase
- Enhanced import organization and code structure

Bug Fixes
- Remove 'ttl' from cache control for Claude Code in request processing
- Simplify system prompt hash calculation in ClaudeCodePreprocess
- Update push paths in dev-build workflow for better file tracking
Features
- Add support for Claude Code and system prompt caching in documentation

Features
- OAuth2 Integration: Added OAuth2 authorization flow to replace PKCE implementation for improved authentication
- Enhanced Caching: Added system prompt hash to ClaudeCodeState and improved cookie management for better caching performance
Improvements
- Configuration Optimization: Enhanced configuration retrieval with improved caching and hash implementation
- Error Handling: Simplified error handling for memberships and organization retrieval in get_organization method
- Code Quality: Refactored OAuth client initialization to remove unnecessary conversions and improve readability
Dependencies
- HTTP Client: Replaced rquest with wreq for HTTP client functionality
- Updated Dependencies: Updated async-compression, toml, toml_parser, toml_writer, and winnow to latest versions
Files Changed
- 22 files modified with 365 insertions and 276 deletions
- Major changes to authentication flow, caching system, and HTTP client implementation

Enhancements
- Improved error handling for rate limits by providing a default reset time of 1 hour when none is specified.
- Enhanced logging for rate limit errors to include the duration until expiration in hours.

Bug Fixes
- Fixed to_oai method to correctly handle ClaudeCodeContext in the response transformation logic, ensuring proper streaming of events in OpenAI API format.

Bug Fixes
- Fixed reset time calculation for rate limit exceeded errors
- Improved handling of missing reset time in rate limit error responses
- Simplified token check logic in check_token method
Refactoring
- Renamed preprocess and context structs for better clarity in Claude middleware
- Moved system message preprocessing to middleware for better organization

🐛 Bug Fixes
- Claude Processing: Fix stop field in Claude Format.

🐛 Bug Fixes
- Claude Processing: Force stop field to be null in Claude Format.

CLAUDE CODE SUPPORT
- Added support for Claude Code API, allowing users to interact with Claude Code API with Cookie authentication.
- Thanks to !!!

Enhancements
- Improve cookie extraction and parsing with better error handling.
Bug Fixes
- Fix gemini inline data field name.

Enhancements
- Update ClewdR name format to include current timestamp for better uniqueness.