Bug: preserve Anthropic thinking replay through proxy-auth translation and clarify prompt-cache behavior

[lidge-jun]

Client or integration

Claude Code

Also reported with Claude Desktop 3P against a remote OpenCodex hub, using a connected client's proxy admission credential.

Area

Proxy and routing

Summary

Separate follow-up for the thinking/replay and prompt-cache request in #3646. This issue applies when Anthropic is the intended upstream. It does not cover Desktop date/hash alias discovery, hub registry ownership, client profile apply, or rejection of unresolved managed aliases; those belong to the remote-alias slice of #3646. Landing that slice does not establish that this replay/cache request is resolved.

The reporter asks that assistant thinking and redacted_thinking blocks survive subsequent turns to the intended Anthropic route, or that a supported untranslated path preserve the original Anthropic exchange. The guide should also explain how proxy authentication, native passthrough eligibility, and prompt-cache retention differ.

Evidence and limits:

• The current inbound translator explicitly drops assistant thinking and redacted_thinking blocks on replay (src/claude/inbound.ts, assistant-content conversion). The Claude Code guide documents this policy.
• Native passthrough requires an eligible unchanged Anthropic model and an actual Anthropic credential; a hub admission token by itself is not that credential. An exposed listener also requires the dedicated proxy admission header. Proxy-auth translation and native passthrough are different routes.
• Translation does not mean prompt caching is disabled. The Anthropic adapter still applies prompt-cache controls, including automatic caching for the native Anthropic endpoint. cacheRetention supports short (default, five minutes), long (one hour), and none.
• Bug: remote-hub Claude Desktop date aliases are forwarded to Anthropic as Opus (quota burn + cache miss) #3646 reports repeated cache misses and faster quota consumption after thinking responses. Neither a per-turn miss guarantee nor its causal attribution to dropped replay blocks has been reproduced in this follow-up. The report's claim that native Claude typically uses one-hour breakpoints also needs a controlled comparison. Do not infer billing or quota savings from the alias fix or from the configured retention duration alone.

Expected outcome:

• Define and implement supported Anthropic-to-Anthropic replay fidelity for proxy-auth traffic, preserving upstream-issued thinking signatures and opaque redacted-thinking blocks where the upstream/client contract requires them. Do not manufacture signatures or forward provider-specific blocks to unrelated providers.
• Verify a multi-turn thinking/tool-result exchange through the intended Anthropic route, including both block types, without replay validation errors or silently dropping required blocks.
• Compare cache creation/read usage across controlled continuation turns with stable model, credential scope, prompt prefix, tools, and retention settings. Distinguish replay changes from TTL expiry and other cache eligibility rules; record unexplained misses as unresolved.
• Document native passthrough eligibility separately from translated-route cache support and the existing retention choices. A global default change to one hour is not implied by this request.

Reproduction

Reported setup from #3646; this follow-up has not executed a new reproduction:

1. Connect Claude Code or Claude Desktop 3P to a remote OpenCodex hub using proxy authentication. Configure the hub's intended route to Anthropic with its existing managed credential.
2. Use a genuine, deliberately selected Anthropic model. Exclude unresolved Desktop aliases and other-provider routes from the comparison.
3. Continue a conversation after an assistant thinking response, including a tool-result continuation. Compare the types and ordering of replayed blocks across the translated boundary using synthetic or privately held content.
4. Record aggregate cache-read/cache-creation usage and elapsed time between turns. Compare supported short/long retention settings under otherwise stable conditions.
5. Separately exercise an upstream-provided redacted-thinking block in a controlled protocol fixture. No such redacted block or valid signed replay fixture was attached to Bug: remote-hub Claude Desktop date aliases are forwarded to Anthropic as Opus (quota burn + cache miss) #3646; current-version coverage remains to be collected.

The existing report's HTTP 404/429 table demonstrates the alias/routing symptoms. It is not a cache-replay reproduction or a measured cache-hit comparison.

Version

Original reporter: OpenCodex 2.42.0 Windows client and 2.41.0 WSL client; Claude Code 2.1.261. The exact remote hub build was not independently established in the issue body. A maintainer comment interpreted the version split differently, so record actual hub/client builds during reproduction rather than treating that interpretation as verified.

Current-version runtime reproduction: pending. This follow-up only confirms the translation policy and cache-control configuration by source inspection.

Operating system

Original report: Windows 11 and WSL, with a remote hub. Exact hub OS/build: not supplied.

Provider and model

Anthropic OAuth pool on the hub; original report names claude-opus-4-8 and claude-haiku-4-5. This follow-up concerns intentional Anthropic routing, independently of GPT/GLM Desktop alias failures.

Logs or error output

No isolated replay failure or controlled cache-usage trace is attached yet. The public report contains alias-related 404/429 observations; do not reuse those as proof of this separate remainder. No request content, credentials, thinking text, signatures, or opaque redacted payloads are included here.

Screenshots and supporting files

• Original request: Bug: remote-hub Claude Desktop date aliases are forwarded to Anthropic as Opus (quota burn + cache miss) #3646
• Maintainer recommendation to split thinking/cache follow-up: Bug: remote-hub Claude Desktop date aliases are forwarded to Anthropic as Opus (quota burn + cache miss) #3646 (comment)
• Existing public guide: https://opencodex.me/guides/claude-code/#native-claude-passthrough-subscription-pierce
• Repository guide: docs-site/src/content/docs/guides/claude-code.md, inbound translation table and native passthrough section.

Duplicate search found no separate issue covering this exact Anthropic proxy-auth replay/cache request:

• [Feature/UX]: add additive Claude Desktop catalog mode with honest native/OpenCodex switching #1213 remains open for honest Desktop native/gateway coexistence and switching. It does not own thinking replay or cache fidelity.
• [Bug]: Intermittent consecutive zero cache hits for Hermes requests through OpenCodeX #3433 remains open for Hermes Chat Completions to native OpenAI/Codex cache affinity. It is a different protocol/provider path.
• Claude Desktop 3P Auto Mode classifier retries after 64-token Anthropic OAuth outputs #545 is closed after stack 4/7: keep an explicit thinking disable through translation (#545) #954 preserved explicit thinking: disabled for the Desktop classifier. That is not assistant thinking-block replay; its correlated classifier sample actually recorded cache hits.
• The request couldn’t be completed. #994 is a closed OpenCode reasoning-content continuation report, with missing provider/model detail at closure. A comment mentions the same inbound drop policy, but it does not cover the specific Anthropic prompt-cache request here.
• [architecture][memory] Make 32 concurrent tool-recall sessions protocol-safe and memory-bounded #820 is a closed concurrency/memory architecture issue with broad protocol-preservation requirements, not an active owner for this remainder.

Search scope: all issue states; title/body/comments for redacted_thinking, thinking cache, thinking replay, cacheRetention, and passthrough proxy; additional Anthropic/cache and open cache searches. Related issue bodies and relevant later comments were inspected, not just titles.

Redacted configuration

Reported configuration shape, not an executable configuration file:

hub runtimeRole: hub
connected client runtimeRole: client
Claude integration authMode: proxy
client credential: proxy admission token only (value omitted)
intended upstream: Anthropic, using a configured hub credential
cacheRetention: short (default); compare long separately

Checks

• I searched existing issues and documentation.
• I removed secrets, tokens, account details, request credentials, and personal data.

[lidge-jun]

리뷰 · 우선순위 58 / 80

이 이슈는 #3646에서 빠져 나온 Anthropic 의도 경로의 thinking/replay·prompt-cache 후속입니다. 지금 dev(HEAD 115b3fc2b)의 src/claude/inbound.ts는 assistant thinking / redacted_thinking 블록을 재생할 때 v1 정책으로 버립니다. 가이드 docs-site/src/content/docs/guides/claude-code.md 번역 표에도 “Dropped”로 적혀 있습니다. 네이티브 Anthropic passthrough는 자격이 되는 그대로의 Anthropic 모델 + 실제 Anthropic 자격 증명이 필요합니다. 허브 입장(admission) 토큰만으로는 passthrough가 켜지지 않습니다. 번역 경로라도 Anthropic 어댑터는 cacheRetention(short 기본 5분 / long 1시간 / none)으로 프롬프트 캐시 컨트롤을 적용합니다. 별칭 수정(#3720)만으로 캐시 히트·할당량 절약이 증명되지는 않습니다.

이슈 본문이 잘 가른 점이다. Desktop 날짜/해시 별칭·허브 레지스트리·클라이언트 apply·미해석 managed alias 거절은 #3646/#3720 쪽이고, 여기는 Anthropic→Anthropic 재생 충실도와 캐시 비교만 남깁니다. 기대 결과도 네 칸으로 분명합니다. (1) 서명·불투명 redacted 블록을 위조하지 않고 필요한 곳에서 보존, (2) thinking+tool-result 다턴 검증, (3) 모델·자격·프리픽스·툴·retention을 고정한 캐시 create/read 비교, (4) passthrough 자격 vs 번역 경로 캐시 vs retention 선택지를 문서에서 분리. 전역 기본을 1시간으로 바꾸라는 요구는 아니라고 못 박았습니다.

중복 검색도 본문에 정리되어 있습니다. #1213 Desktop native/gateway 공존, #3433 Hermes/OpenAI 캐시 친화, 닫힌 #545/#994/#820은 이 잔여의 주인이 아닙니다. 현재 버전 런타임 재현은 pending이고, 소스 검사로 번역 정책과 캐시 설정만 확인한 상태입니다. #3720이 #3646 별칭을 닫으면 이 이슈가 남은 Anthropic 재생 작업의 정식 추적 티켓이 됩니다.

우선순위 58인 이유다. 계약·할당량에 닿는 실버그 후속이지만, 재현 측정이 아직 없고 별칭 랜딩(#3720) 뒤에 설계가 더 선명해집니다. 지금 당장 코드를 급히 넣기보다 스펙·픽스처를 잠그는 단계가 맞습니다.

경로/심볼 - src/claude/inbound.ts case "thinking" / case "redacted_thinking" - 재생 시 break로 드롭(v1).
경로 - docs-site/src/content/docs/guides/claude-code.md inbound 표 - thinking/redacted_thinking replay = Dropped; native passthrough 절과 분리 필요.
경로/심볼 - src/adapters/anthropic.ts createAnthropicAdapter / resolveCacheControl - 번역 경로에서도 cacheRetention 적용.
경로/심볼 - src/types/config.ts cacheRetention?: "none"|"short"|"long" - 기본 short; 글로벌 long 강제 없음.
경로 - #3646 / PR #3720 - 별칭 조각과 이 이슈를 섞지 말 것(본문·#3720 요약과 일치).
경로 - #1213 #3433 - 관련이나 소유권 이전 대상 아님.

메인테이너의 판단이 필요한 지점

• Anthropic→Anthropic에서 블록을 그대로 통과시킬지, 서명 검증 가능한 최소 보존만 할지
• proxy-auth 번역 경로에 “지원되는 미번역 우회”를 새 모드로 둘지, 기존 native passthrough 자격만 강화할지
• 캐시 비교를 CI 픽스처로 강제할지, 유지보수자 수동 측정 체크리스트로 둘지
• #3720 머지 전에 이 이슈 구현을 시작할지, 별칭 안정화 후 할지

너의 추천
이슈를 열어 두고 #3720 랜딩 뒤에 구현 설계를 잡으세요. 라벨 bug/proxy는 유지. 첫 PR은 inbound drop 완화 + 다턴 픽스처 + 가이드 문구 분리까지를 한 사이클로, 캐시 계측은 두 번째 사이클로 나누는 편이 안전합니다. 중복 이슈로 닫지 마세요. #3646 close 시 이 이슈 링크만 남기면 됩니다.

이 댓글은 grok-bot이 작성했습니다