feat(anthropic): use official Anthropic endpoints/models with API+SDK routes#107
Draft
azkore wants to merge 19 commits intoSoju06:mainfrom
Draft
feat(anthropic): use official Anthropic endpoints/models with API+SDK routes#107azkore wants to merge 19 commits intoSoju06:mainfrom
azkore wants to merge 19 commits intoSoju06:mainfrom
Conversation
Add a dedicated `/claude/v1/messages` Anthropic-compatible endpoint backed by the local Claude SDK runtime while keeping existing OpenAI routes active in the same server. Wire Anthropic request logs, pricing, and usage windows into existing dashboard metrics, including OAuth usage polling and corrected cached input token accounting for Claude usage payloads.
Reduce Claude SDK connect/disconnect churn by reusing connected clients keyed by request options. Add pool settings and close all pooled clients during app shutdown to prevent leaked sessions.
Serve Anthropic-compatible messages over both the existing SDK-backed /claude endpoint and a new direct /claude-api endpoint so clients can choose runtime parity needs without changing servers. Add OAuth refresh-aware API transport plumbing, configuration docs, and integration coverage for both non-streaming and streaming paths.
Exclude only the generated Anthropic provider seed account from the "already logged in" browser shortcut so OAuth can still fall back to device flow when no real account is present. Add integration coverage for both seed-account fallback and non-seed anthropic_default account handling.
Add a dedicated Anthropic credential import path in the dashboard so Claude OAuth JSON can be uploaded without restarts and used by the /claude-api transport immediately. Prefix Anthropic account display names with claude/ while leaving OpenAI account handling unchanged, and add backend/frontend coverage.
Require Claude credential imports to provide an explicit account email instead of inferring identity from credential payloads that often omit user metadata. Update the dashboard import dialog and API contract so claude/ labels use user-provided email deterministically.
Add request-shape diagnostics for both Anthropic transports so payload size, mutation, and usage data can be correlated during token blow-up investigation. Harden SDK pooling for stateful usage by keying pools by session id and keeping ephemeral session traffic unpooled to avoid context bleed.
Retry Anthropic API requests after token refresh not only on 401, but also on revoked-token 403 responses returned by OAuth-backed endpoints. This preserves automatic recovery when access tokens are revoked but a valid refresh token is still available.
Exclude Anthropic provider accounts from the OpenAI usage refresh loop so Claude tokens are not sent to /backend-api/wham/usage and noisy 401 errors are avoided. Keep Anthropic usage refresh on its dedicated scheduler path unchanged.
Turn off Anthropic SDK client pooling by default to avoid state bleed in stateful conversation workloads until pooling behavior is fully hardened. Before re-enabling pooling, fix all of the following: - keep strict client affinity per explicit session id - prevent client reuse for ephemeral/default session traffic - serialize concurrent access per pooled session client - detect/reset broken pooled clients after stream/query failures - add regression coverage for cross-session context contamination - add pool observability for hit/miss, reuse, and session binding
Switch the default Anthropic route to the direct OAuth API transport at /claude/v1/messages and move the Claude SDK transport to /claude-sdk/v1/messages. Update docs, examples, and integration coverage so route semantics are explicit and consistent for testing and rollout.
️✅ There are no secrets present in this pull request anymore.If these secrets were true positive and are still valid, we highly recommend you to revoke them. 🦉 GitGuardian detects secrets in your source code to help developers and security teams secure the modern development process. You are seeing this because you or someone else with access to this repository has authorized GitGuardian to scan your pull request. |
Fix strict type-checker failures in Anthropic diagnostics/usage paths so Ruff and ty pass in CI. Replace token-like test fixture literals with clearly fake values to avoid secret scanner false positives.
Contributor
Author
|
the GitGuardian findings were from test fixture strings (not real credentials). I replaced the token-like test values and rewrote this branch so the previously flagged literals are no longer present in this PR history. |
Apply Ruff formatter output to files flagged by CI format check so the Lint (ruff) workflow passes on the rewritten branch history.
Stop prefixing Anthropic display names with claude/ so account labels remain clean and consistent with existing email presentation. Add provider badges, subtle Anthropic row tinting, and provider-grouped sorting in account/request views for clearer visual separation.
Replace textual provider badges with compact provider icons and extend Anthropic tinting to account cards/details for clearer visual separation. Remove internal account-id disambiguation from account subtitles and use provider labeling instead so user-facing lists stay clean.
Add focused tests for account provider helper branches so Vitest coverage clears the global branch threshold again. Also keep accounts mapper formatting aligned with Ruff to avoid format-check failures in CI.
Align Claude pricing constants with current Anthropic pricing for Opus 4.6/4.5/4.1, Sonnet 4.6/4.5, and Haiku 4.5. Also expand alias coverage for hyphenated and provider-prefixed model names so cost estimates resolve to the correct tier.
codex-lb logs and dashboard views are OpenAI-focused, where input tokens represent full request context for comparison and cost views. Anthropic reports input as separate parts (base input, cache creation, cache read). Log Anthropic input as the sum of those parts so Anthropic rows follow the same convention and are comparable in the UI. Update SDK and API usage parsing tests for the new accounting rule.
29 tasks
Contributor
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this PR does
This PR makes it possible to use official Anthropic endpoints and models through
codex-lb, with two transport options:POST /claude/v1/messages-> official Anthropic API (default)POST /claude-sdk/v1/messages-> local Claude SDK/runtimeIn simple terms:
codex-lbcan now use official Anthropic endpoints/models in two transport modes, and the default/clauderoute uses the official Anthropic API path.How this differs from other open PRs
This PR is not a duplicate of current open Anthropic/Claude PRs:
This PR focuses on using official Anthropic endpoints/models via dedicated Anthropic routes.
Main changes
/claude/v1/messages/claude-sdk/v1/messages/api/accounts/import-anthropic) with explicit email input.claude/<email>while keeping existing Codex/OpenAI account behavior unchanged.Anthropic API vs Anthropic SDK
/claude/v1/messages(API route): direct call toapi.anthropic.com/v1/messageswith OAuth credentials, with optional CLI-header/system-prompt parity helpers./claude-sdk/v1/messages(SDK route): localclaude-agent-sdktransport via Claude runtime session.Attribution
This implementation is heavily inspired by ccproxy-api, especially:
ccproxy/plugins/claude_api/*ccproxy/plugins/oauth_claude/*The ideas were adapted to
codex-lbarchitecture, routes, and dashboard/account flow.Current limitations
CODEX_LB_ANTHROPIC_SDK_POOL_ENABLED=false).Pooling checklist before re-enable
Testing
uv run pytest).Collaboration note
I am testing this branch now, but I may not have time to continue iterating quickly.
Anyone who wants to test, review, or continue work on this draft PR is very welcome.