feat(issue-37): multi-provider subscription engine (Codex, Gemini, Qwen) + token refresh#38
Conversation
Adding .gitkeep for PR creation (default mode). This file will be removed when the task is complete. Issue: #37
Adopt the best experience from heyhuynhgiabuu/proxypal to plan full support for Claude, Codex, Gemini, and Qwen subscriptions. - README.md: analysis + gap matrix (ProxyPal is a UX shell over CLIProxyAPI) - requirements.md: process + functional requirements trace - solution-plans.md: file-level phased plan per requirement + execution order - components-survey.md: build-vs-borrow survey (CLIProxyAPI, oauth2 crate, ...) - online-research.md: cited per-provider OAuth endpoints/tokens/quotas - proxypal-analysis.md: deep inventory of ProxyPal + CLIProxyAPI engine - raw/: issue JSON + ProxyPal snapshot metadata Remove placeholder .gitkeep now that real content exists. Refs #37
konard
changed the title
The v0.19.0 release (ac24fe4) bumped Cargo.toml to 0.19.0 but left Cargo.lock's link-assistant-router entry at 0.18.0. 'cargo package --list' re-syncs the lock and then fails on the resulting dirty working tree, breaking the Build Package CI job. Sync the lock to match Cargo.toml.
🤖 Solution Draft LogThis log file contains the complete execution trace of the AI solution draft process. 💰 Cost estimation:
📊 Context and tokens usage:Claude Opus 4.8: (2 sub-sessions)
Total: (253.1K new + 1.6M cache writes + 7.1M cache reads) input tokens, 271.5K output tokens, $21.651696 cost Claude Haiku 4.5:
Total: 1.3M input tokens, 52.6K output tokens, $2.355329 cost 🤖 Models used:
📎 Log file uploaded as Gist (6824KB)Now working session is ended, feel free to review and add any feedback on the solution draft. |
✅ Ready to mergeThis pull request is now ready to be merged:
Monitored by hive-mind with --auto-restart-until-mergeable flag |
|
We need implement all the features of underlying tech to support subscriptions, API routing and so on. With no UI support. Only the best practices for the cli/docker/library. And we also need to update all our dependencies to latest versions (using latest stable Rust only). |
|
🤖 AI Work Session Started Starting automated work session at 2026-06-17T17:53:47.507Z The PR has been converted to draft mode while work is in progress. This comment marks the beginning of an AI work session. Please wait for the session to finish, and provide your feedback. |
- Bump edition 2021 -> 2024, rust-version 1.70 -> 1.85 (latest stable Rust) - Major dependency bumps: reqwest 0.12->0.13, jsonwebtoken 9->10, tower-http 0.6->0.7, log-lazy 0.1->0.2, sha2 0.10->0.11; minor bumps for hyper, tokio, bytes, http and others via cargo update - jsonwebtoken 10 requires explicit crypto provider: enable rust_crypto (pure-Rust, no aws-lc C deps) to keep the single-binary build portable - Reformat imports to edition 2024 rustfmt style - Replace io::Error::new(ErrorKind::Other, _) with io::Error::other(_) All tests green (82 + 49 + 10); clippy clean; fmt clean.
Introduce src/subscription.rs: a generalized, fully unit-tested reader for Claude/Codex/Gemini/Qwen subscription OAuth tokens cached by their vendor CLIs. Normalizes the four on-disk layouts (Claude nested claudeAiOauth, Codex tokens block + chatgpt-account-id, Gemini/Qwen flat oauth_creds.json with expiry_date/resource_url) into a single SubscriptionToken with base-URL resolution. Foundation for F2/F3/F4/F5.
…eams
Wire the SubscriptionReader into AppState and dispatch chat-completions,
responses, and model-listing routes to the right vendor backend:
- Codex (ChatGPT): translate Chat Completions -> Responses API and forward
to the ChatGPT backend with chatgpt-account-id / openai-beta headers.
- Qwen: forward OpenAI-compatible requests to DashScope, honoring the
per-account resource_url base.
- Gemini: translate to/from the Code Assist generateContent dialect,
synthesizing SSE when the client requested streaming.
Adds UpstreamProvider::{Codex,Gemini,Qwen} variants plus
subscription_provider() mapping, openai::chat_completion_to_responses,
src/gemini.rs, and src/subscription_proxy.rs.
Extend `router doctor` to report each vendor subscription credential file: whether it is present, and whether the token is readable, valid, or expired.
When an on-disk Codex/Gemini/Qwen token has expired, exchange its refresh_token for a fresh access token using each vendor's public OAuth client and cache the result in memory. Vendor credential files stay read-only; secrets are never logged. Wired into both the OpenAI-shaped subscription forwarder and the Gemini Code Assist path via a shared TokenCache on AppState. This keeps the proxy working even when the vendor CLI is not running to refresh its own credential file (the key ProxyPal behavior).
Forward Retry-After and x-ratelimit-* headers from subscription upstreams to the client on both streaming and buffered responses, so callers can honor provider throttling and back off.
Add a changelog fragment (minor bump) and README coverage: the new UPSTREAM_PROVIDER values, credential locations, in-memory refresh, and dialect translation for the vendor subscription upstreams.
Mark F2-F5/F7/F11 as Done (subscription reading, routing, dialect translation, in-memory refresh), F8/F10 as Partial, and F6/F9/F12 as deferred/out-of-scope per the no-UI governing directive.
konard
changed the title
src/openai.rs exceeded the repo's 1000-line file cap after the Codex Responses-API translation was added. Move the Responses-API request/response translators (OpenAIResponseRequest, response_to_anthropic, chat_completion_to_responses, anthropic_to_response) into a dedicated src/responses.rs module, sharing the field-shaping helpers (extract_text, translate_tools) via pub(crate). No behaviour change.
Working session summaryAll tasks are complete. Here's the final status. SummaryThe blocker — a leaked Google OAuth client secret committed in What I did this session
Final verification
PR: #38 This summary was automatically extracted from the AI working session output. |
🤖 Solution Draft LogThis log file contains the complete execution trace of the AI solution draft process. 💰 Cost estimation:
📊 Context and tokens usage:Claude Opus 4.8: (5 sub-sessions)
Total: (93.8K new + 510.8K cache writes + 17.0M cache reads) input tokens, 160.6K output tokens, $16.175775 cost 🤖 Models used:
📎 Log file uploaded as Gist (7692KB)Now working session is ended, feel free to review and add any feedback on the solution draft. |
✅ Ready to mergeThis pull request is now ready to be merged:
Monitored by hive-mind with --auto-restart-until-mergeable flag |
1 participant
What this PR does
Implements multi-provider subscription support and API routing in the router
engine, adopting the best practices from
heyhuynhgiabuu/proxypal /
CLIProxyAPI, so the router can
"fully support claude, codex, gemini, qwen, and their subscriptions with all our
features and more."
Per the governing directive on this PR, this is implemented as a CLI / Docker /
library engine with no UI, on the latest stable Rust with all
dependencies updated.
Engine changes
src/subscription.rs) — reads eachvendor CLI's credential file read-only:
~/.codex/auth.json,~/.gemini/oauth_creds.json,~/.qwen/oauth_creds.json(Claude via theexisting
~/.claudereader). The router never writes to vendor locations.src/subscription_proxy.rs,src/gemini.rs) — newUpstreamProvider::{Codex, Gemini, Qwen}upstreams route/v1/chat/completions,/v1/responses, and/v1/modelsto the correctvendor backend (ChatGPT Codex Responses API, Gemini Code Assist
generateContent,DashScope OpenAI-compatible).
(Codex) ↔ Gemini Code Assist
generateContent, including SSE synthesis whena client requests streaming from Gemini.
src/refresh.rs) — expired Codex/Gemini/Qwentokens are refreshed via each vendor's public OAuth client and cached in
memory only; vendor files stay read-only and secrets are never logged. No
OAuth client secret is hardcoded — Gemini's installed-app secret is supplied at
runtime via the
GEMINI_OAUTH_CLIENT_SECRETenv var.router doctornow probes each subscription credential file and reportspresent / token OK / EXPIRED / NO TOKEN / MISSING.
Retry-Afterandx-ratelimit-*headers fromsubscription upstreams are passed through so clients can back off intelligently.
Dependencies & toolchain
latest stable toolchain (
build(deps): ...commit).Case-study package
The evidence-backed analysis under
docs/case-studies/issue-37/collects the data, analyzes ProxyPal/CLIProxyAPI, lists every requirement
(
requirements.md, F1–F12), proposes a plan per requirement(
solution-plans.md), and surveys components (components-survey.md). Onlineresearch with cited primary sources and redacted secrets is in
online-research.md.Key finding
ProxyPal is a UX shell, not an engine — a Tauri/SolidJS GUI wrapping the Go
CLIProxyAPI binary. "Adopt the best of ProxyPal" therefore resolves to adopting
CLIProxyAPI's engine architecture (per-provider OAuth, dialect translation,
account rotation/cooldown) natively in our Rust engine — which this PR does — while
its GUI/auto-config UX is explicitly out of scope per the no-UI directive.
Requirement status (functional, F1–F12)
Full trace in
requirements.md.Security
~/.codex,~/.gemini,~/.qwen,~/.claude.unsafe_code = "forbid"; clippy all/pedantic/nursery clean.Testing
cargo clippy --all-targets --all-features→ 0 warnings.cargo test→ 170 tests pass (111 lib + 49 integration + 10 release-tooling),including unit tests for the refresh layer, rate-limit header selection, and URL
joining per provider.
changelog.d/fragment (bump: minor) documents the new capabilities.Closes #37