feat(proxy): add 200+error body detection and extra inputs rectifier by NKU100 · Pull Request #2433 · farion1231/cc-switch

NKU100 · 2026-04-28T22:14:28Z

Summary

Add extra inputs rectifier module to auto-strip unsupported API fields, with per-provider caching (1-hour TTL). Detects "extra inputs are not permitted" errors and falls back to checking request body for known Anthropic-only fields (context_management, anthropic_beta, output_config) when error messages don't specify which fields are unsupported
Detect non-streaming 200 responses that contain an error key and reclassify them as UpstreamError, enabling rectifiers and failover to trigger
Decompress gzip response body before validity check in hyper path (handles proxies that strip gzip wrapper but leave content-encoding header)
Skip gzip decompression when magic bytes are absent (prevents ZlibError from misleading content-encoding headers)

Test plan

cargo clippy -- -D warnings passes
cargo fmt --check passes
All unit tests pass (1053+ tests)
Verified live with Xiaomi MiFE API returning gzip-compressed 200 responses
Verified live with a third-party provider returning 200 + {"error":...} — error-key detection → extra inputs rectifier chain triggers correctly
CI passes on GitHub Actions

… fields When a provider returns "Extra inputs are not permitted" errors, the rectifier extracts the unsupported field names from the error message, caches them per-provider (1-hour TTL), strips the fields, retries the request, and pre-filters subsequent requests from the same provider. Includes settings panel toggle (en/zh/ja) and retry guard to prevent infinite loops.

Remove unused `extracted_fields` field and `insert` method, use char array pattern in `rsplit` to satisfy clippy::manual_pattern_char_comparison.

…elds Non-streaming 200 responses lacking expected keys (content/choices/output/ candidates) are now reclassified as UpstreamError, triggering rectifiers and failover. When the extra-inputs rectifier fires but cannot extract field names from the error message, it falls back to stripping known Anthropic-only fields (context_management, anthropic_beta, output_config) present in the request body.

farion1231 · 2026-04-29T01:47:22Z

@codex review

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5b68a192c0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-04-29T01:52:45Z

+    obj.contains_key("content")
+        || obj.contains_key("choices")
+        || obj.contains_key("output")
+        || obj.contains_key("candidates")


Accept prompt-blocked Gemini responses as valid

is_valid_response_body only treats objects containing content/choices/output/candidates as success, so a valid Gemini generateContent 200 response that only carries promptFeedback (blocked prompt, no candidates) is reclassified as UpstreamError. In that case this code triggers unnecessary retry/failover instead of returning the provider’s safety-block result to callers, which changes behavior for normal filtered prompts.

Useful? React with 👍 / 👎.

这个不太好修啊，不知道补上generateContent能不能覆盖所有场景

chatgpt-codex-connector · 2026-04-29T01:52:45Z

+        let entries = self.entries.read().await;
+        entries
+            .iter()
+            .filter(|(k, v)| k.starts_with(&prefix) && !v.is_expired())
+            .map(|(k, _)| k[prefix.len()..].to_string())
+            .collect()


Evict expired cache entries when reading blocked fields

get_blocked_fields filters out expired entries but never removes them from entries, so the cache grows monotonically in long-running processes as new provider/field pairs are learned. This violates the intended 1-hour TTL behavior in practice (expiration affects matching, not retention) and can cause avoidable memory growth over time.

Useful? React with 👍 / 👎.

已经在最新提交修复了

…ache entries - Add "promptFeedback" to is_valid_response_body so Gemini safety- blocked responses are forwarded instead of misclassified as errors - Lazily evict expired entries in ExtraInputsCache::get_blocked_fields to prevent unbounded memory growth in long-running processes

Replace the key-presence allowlist in is_valid_response_body with discriminator-field matching per wire format: - Anthropic Messages: type == "message" - OpenAI Chat Completions: object == "chat.completion" - OpenAI Responses API: object == "response" - Gemini: candidates / promptFeedback (no discriminator in protocol) This avoids misclassifying 200 bodies that happen to contain strings like "content" or "choices" as errors, and correctly rejects error envelopes (type == "error") and SSE chunk types.

farion1231 · 2026-05-01T14:12:09Z

@codex review

chatgpt-codex-connector · 2026-05-01T14:19:33Z

Codex Review: Didn't find any major issues. Already looking forward to the next diff.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

farion1231 · 2026-05-02T08:22:35Z

@claude review

claude · 2026-05-02T08:22:50Z

Claude Code is working…

I'll analyze this and get back to you.

View job run

farion1231 · 2026-05-02T08:24:11Z

@claude review

(Retrying — first attempt hit a git fetch race in claude-code-action.)

claude · 2026-05-02T08:24:28Z

Claude Code is working…

I'll analyze this and get back to you.

View job run

farion1231 · 2026-05-02T11:28:37Z

@claude review

claude · 2026-05-02T11:28:57Z

Claude Code is working…

I'll analyze this and get back to you.

View job run

farion1231 · 2026-05-02T12:34:36Z

@claude review

claude · 2026-05-02T12:34:49Z

Claude finished @farion1231's task in 4m 50s —— View job

Claude PR Review

Read PR diff
Review extra_inputs_rectifier.rs
Review forwarder.rs (200+error body detection)
Verify Codex review concerns are addressed
Post review

Findings

Important — is_valid_response_body will reject valid Gemini :countTokens / :embedContent 200 responses (confidence: 85)

forwarder.rs:1687-1712 runs is_valid_response_body on every non-SSE 200 response. The discriminator-based check accepts only Anthropic (type=="message"), OpenAI (object=="chat.completion"|"response"), or Gemini bodies that contain candidates/promptFeedback.

But the Gemini route in server.rs:328 is a wildcard:

.route("/v1beta/*path", post(handlers::handle_gemini))
.route("/gemini/v1beta/*path", post(handlers::handle_gemini))

So it also forwards non-generateContent Gemini endpoints — and their successful responses do not contain any of the keys above:

:countTokens → {"totalTokens": 1234, "totalBillableCharacters": 567}
:embedContent → {"embedding": {"values": [...]}}
:batchEmbedContents → {"embeddings": [...]}

Concrete failure path: Gemini CLI (and similar clients) call :countTokens before each generateContent. cc-switch receives the 200 from upstream, is_valid_response_body returns false, forward() raises ProxyError::UpstreamError { status: 200, ... }, the request is retried against every configured provider — all of which return the same shape — and the client eventually sees a hard failure on what was a perfectly valid response.

Two minimal-surface fixes:

Skip the discriminator check when the endpoint is not a generation endpoint (e.g. endpoint.contains("generateContent") || endpoint.contains("/messages") || endpoint.contains("/chat/completions") || endpoint.contains("/responses")), or
Extend the Gemini branch in is_valid_response_body to also accept totalTokens, embedding, or embeddings as keys.

Fix this →

Notes on prior Codex feedback

Both items are already resolved on this branch, looks good:

Gemini promptFeedback-only 200 is accepted (commit f7057aa, then refactored in 3a20666) — covered by valid_gemini_safety_blocked_response test.
ExtraInputsCache now evicts expired entries on read (extra_inputs_rectifier.rs:94-100).

Otherwise the rectifier module is well-scoped, the per-provider keying is correct, the pre-filter ordering sits cleanly before the existing body filter, and the discriminator-based body classifier is a nice tightening over the original key-set check.

…per path The hyper client does not auto-decompress responses, so gzip-compressed 200 bodies (e.g. from Xiaomi mimo API) were misidentified as invalid by is_valid_response_body, triggering unnecessary retries. Add manual decompression using response_processor::decompress_body before the validity check, matching what the reqwest path already does.

…ection Instead of whitelisting valid response shapes per protocol (Anthropic type=="message", OpenAI object=="chat.completion", Gemini candidates/ promptFeedback, ...), detect the `error` key in JSON body. Any 200 response containing `error` is reclassified as UpstreamError; everything else is trusted. This eliminates the need for per-protocol patches when new endpoints or providers are added.

MiFE proxy may strip gzip wrapping but leave content-encoding: gzip header intact. decompress_body then fails with "invalid gzip header", and the misleading header is forwarded to clients causing ZlibError. Check for gzip magic bytes (1f 8b) before attempting decompression. If absent, return the body as-is with Ok, which triggers content-encoding header stripping in read_decoded_body.

…back strategy The old description only mentioned matching "Extra inputs are not permitted" errors. The actual implementation also falls back to checking Anthropic-specific fields (context_management, anthropic_beta, output_config) in the request body. Update zh/en/ja descriptions to reflect this broader detection behavior.

NKU100 · 2026-05-02T17:36:21Z

Updated PR description to reflect the latest implementation changes since the original submission:

200+error detection: Simplified from protocol-specific whitelist (content/choices/output/candidates) to error key detection — no more per-protocol patches needed
Gzip decompression: Added magic bytes guard (1f 8b) to handle proxies (e.g. Xiaomi MiFE) that strip gzip wrapper but leave content-encoding: gzip header, preventing ZlibError on downstream clients
Extra Inputs rectifier description: Updated i18n text (zh/en/ja) to reflect the Anthropic-only fields fallback strategy

NKU100 added 3 commits April 28, 2026 23:10

fix(proxy): resolve clippy dead-code errors in extra inputs rectifier

afacf0b

Remove unused `extracted_fields` field and `insert` method, use char array pattern in `rsplit` to satisfy clippy::manual_pattern_char_comparison.

chatgpt-codex-connector Bot reviewed Apr 29, 2026

View reviewed changes

NKU100 added 2 commits April 29, 2026 11:24

NKU100 added 2 commits May 3, 2026 00:36

NKU100 force-pushed the main branch from 4a9d1e7 to 49bc7d6 Compare May 2, 2026 16:37

NKU100 added 2 commits May 3, 2026 01:02

Uh oh!

Conversation

NKU100 commented Apr 28, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Test plan

Uh oh!

farion1231 commented Apr 29, 2026

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

NKU100 Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

chatgpt-codex-connector Bot Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

NKU100 Apr 29, 2026

Choose a reason for hiding this comment

Uh oh!

farion1231 commented May 1, 2026

Uh oh!

chatgpt-codex-connector Bot commented May 1, 2026

Uh oh!

farion1231 commented May 2, 2026

Uh oh!

claude Bot commented May 2, 2026

Uh oh!

farion1231 commented May 2, 2026

Uh oh!

claude Bot commented May 2, 2026

Uh oh!

farion1231 commented May 2, 2026

Uh oh!

claude Bot commented May 2, 2026

Uh oh!

farion1231 commented May 2, 2026

Uh oh!

claude Bot commented May 2, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Claude PR Review

Findings

Notes on prior Codex feedback

Uh oh!

NKU100 commented May 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

NKU100 commented Apr 28, 2026 •

edited

Loading

claude Bot commented May 2, 2026 •

edited

Loading