Keep Anthropic request body rewrites attempt-local and synchronize the accepted wire body only after upstream success so failover and retry paths do not reuse stale parsed state.
Keep large gateway payloads as raw body ranges and bind OpenAI parsed-body caches to the body bytes so failover and mapping do not reuse stale mutable state.
Anthropic count_tokens rejects generation-only fields such as temperature, top_p, top_k, stream, and stop sequences. Passing the original messages payload through unchanged can turn otherwise valid requests into upstream 400 errors.
Sanitize only the count_tokens upstream payload after the gateway's existing request normalization, preserving fields that existing compatibility paths rely on while removing parameters the count_tokens endpoint does not accept.
Fixes#2764
Co-authored-by: Cursor <cursoragent@cursor.com>
Pool mode currently retries the same account for a fixed set of
upstream HTTP statuses: 401, 403, 429. Some upstream pool deployments
also need same-account retry for transient provider/proxy statuses
such as 502, 503, 520, 529, but hard-coding more statuses changes
behavior for everyone.
Add a per-account credentials option `pool_mode_retry_status_codes`
that lets admins choose which upstream HTTP status codes trigger
same-account retry in pool mode:
- Unset (default): preserve the current 401/403/429 default
- Explicit list: override the defaults with the configured codes
- Codes normalized to the 100-599 range, deduplicated, sorted
The standalone `isPoolModeRetryableStatus` helper is kept as the
default-only fallback. All 15 gateway call sites switch to the new
`Account.IsPoolModeRetryableStatus` method so behavior is preserved
for accounts that do not configure the new field.
Frontend admin UI gains a "Retry Status Codes" comma-separated input
under the pool-mode section in both Create/Edit account modals
(en + zh i18n).
Fixes#2731
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Export ApplyBedrockCCCompat() in GatewayService, called after channel
model mapping to ensure mapped model ID is used for Opus 4.7+ detection
- Add sanitizeBedrockCCFields(): remove service_tier/interface_geo/
context_management, inject max_tokens/anthropic_version defaults
- Add sanitizeBedrockCCBetaTokens(): filter anthropic_beta to keep only
Bedrock-supported tokens, reusing autoInjectBedrockBetaTokens and
filterBedrockBetaTokens for consistent rules
- Remove unsupported beta tokens (interleaved-thinking, context-management)
from whitelist based on AWS official docs
- Simplify IsBedrockCCCompatEnabled() to check boolean toggle directly,
applying CC compat to all accounts regardless of platform
- Add unit tests for IsBedrockCCCompatEnabled (8 cases),
sanitizeBedrockCCFields (8 cases), sanitizeBedrockCCBetaTokens (7 cases)
- Update bedrock beta policy tests for removed auto-injection
Add channel-level Bedrock CC compatibility toggle (similar to web_search_emulation)
that fixes 4 types of Bedrock 400 errors seen with Claude Code:
1. thinking.type "enabled" → "adaptive" for Opus 4.7+ (only supports adaptive)
2. Add default budget_tokens when missing for older models
3. Replace illegal characters in tool_use IDs to match Bedrock's ^[a-zA-Z0-9_-]+$ pattern
4. anthropic_version / invalid beta flag (already handled elsewhere)
Transformations run in Forward() before any forwarding path, so both native Bedrock
accounts and apikey passthrough accounts pointing to Bedrock relays benefit.
Includes channel-level toggle UI and unit tests.
Some clients reuse assistant history from other models when switching to
claude with extended thinking enabled. If a prior thinking block lacks the
thinking text field, upstream returns:
messages.X.content.Y.thinking: each thinking block must contain thinking
Add this pattern to isThinkingBlockSignatureError so the existing
FilterThinkingBlocksForRetry retry path triggers and rewrites/drops the
offending blocks.
(*net.OpError).Error() concatenates Source/Addr fields, so the previous
disconnectMsg surfaced internal source IP/port and upstream server address
to clients via SSE error frames and UpstreamFailoverError.ResponseBody
(reported by @Wei-Shaw on PR #2066).
- Add sanitizeStreamError that maps known errors (io.ErrUnexpectedEOF,
context.Canceled, syscall.ECONNRESET/EPIPE/ETIMEDOUT/...) to fixed
descriptions and falls back to a generic placeholder, with an explicit
*net.OpError branch that drops Source/Addr fields entirely.
- Use sanitized message in client-facing disconnectMsg; full ev.err is
still preserved in the existing operator log line for diagnosis.
- Tests cover net.OpError redaction, the failover ResponseBody path, and
every known sanitized error mapping.
Two follow-ups to PR #2066's failover-wrap fix:
1. Failover ResponseBody (`UpstreamFailoverError.ResponseBody`) was encoded
as `{"error": "<msg>"}` (string field). `ExtractUpstreamErrorMessage`
probes for `error.message`, `detail`, or top-level `message` only — so
`handleFailoverExhausted` and downstream passthrough rules saw an empty
message, losing the EOF root cause in ops logs. Re-encode as the
Anthropic standard shape `{"type":"error","error":{"type":"upstream_disconnected","message":"..."}}`.
(Addresses the inline review comment from copilot-pull-request-reviewer
on Wei-Shaw/sub2api#2066.)
2. The streaming `event: error` SSE frame for `response_too_large`,
`stream_read_error`, and `stream_timeout` was non-standard
(`{"error":"<reason>"}`). Anthropic SDKs (and Claude Code) expect
`{"type":"error","error":{"type":"...","message":"..."}}` and parse
`error.type`/`error.message` accordingly. Refactor `sendErrorEvent` to
take both reason and message, and emit the standard frame so client
SDKs surface a real diagnostic message instead of a generic stream error.
This does not by itself prevent task interruption on long-stream EOF
(SSE has no resume; client-side retry remains the only complete fix), but
it gives both server-side ops logs and client-side error UIs a meaningful
upstream message so users know the next step is to retry.
Tests updated to assert the new body shape on both branches plus a new
assertion that `ExtractUpstreamErrorMessage` returns a non-empty string.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Anthropic streaming path (gateway_service.go) returned a plain error on
upstream SSE read failure, so the handler-level UpstreamFailoverError check
never fired and the client received a bare `stream_read_error` event,
breaking long-running tasks even when no bytes had been written yet.
The most common trigger is HTTP/2 GOAWAY from api.anthropic.com edge
backends doing graceful rotation: Go's http.Transport surfaces this as
`unexpected EOF` and never auto-retries.
Mirror what the OpenAI and antigravity gateways already do: when the read
error happens before any byte has reached the client (`!c.Writer.Written()`),
return `*UpstreamFailoverError{StatusCode: 502, RetryableOnSameAccount: true}`
so the handler can retry on the same or another account. After client
output has begun, SSE has no resume protocol — keep the existing passthrough
behavior.
Tests cover both branches via streamReadCloser-based fixtures.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PR #1914 unconditionally applied the full mimicry pipeline to all OAuth
accounts, including real Claude Code CLI clients. This replaced the
client's long system prompt (~10K+ tokens with stable cache_control
breakpoints) with a short ~45 token [billing, CC prompt] pair, which
falls below Anthropic's 1024-token minimum cacheable prefix threshold.
The result: every request created a new cache but never hit an existing
one.
Fix: restore the Claude Code client detection gate so that real CC
clients bypass body-level mimicry (system rewrite, message cache
management, tool name obfuscation). Non-CC third-party clients
(opencode, etc.) continue to receive full mimicry.
Also harden the detection logic:
- Make UA regex case-insensitive (align with claude_code_validator.go)
- Validate metadata.user_id format via ParseMetadataUserID() instead of
just checking non-empty, preventing third-party tools from spoofing
a claude-cli/* UA with an arbitrary user_id string to bypass mimicry
- staticcheck QF1001: apply De Morgan's law to the OAuth-mimic header
passthrough guard (`!(a && b)` → `a != ... || !b`).
- unused: drop `isClaudeCodeRequest`, which became dead after PR #1914
switched both `/v1/messages` and `/count_tokens` paths to unconditional
`account.IsOAuth()` mimicry. The lowercase helper `isClaudeCodeClient`
is kept (still referenced by `TestIsClaudeCodeClient`).
Root cause of persistent third-party detection: sub2api's
buildUpstreamRequest transparently forwards client headers via
allowedHeaders whitelist (addHeaderRaw) before applying mimicry
overrides. When third-party clients (opencode, etc.) send their own
anthropic-beta / user-agent / x-stainless-* / x-claude-code-session-id
values, these get appended to the request alongside our injected
headers, creating an inconsistent header set that Anthropic detects.
Parrot's build_upstream_headers constructs exactly 9 headers from
scratch and never forwards anything from the client. This is why
'same opencode version, some users work some don't' — different
opencode configs/versions send different header combinations.
Fix: when tokenType=oauth and mimicClaudeCode=true, skip the
client header passthrough loop entirely. The subsequent
applyClaudeCodeMimicHeaders + ApplyFingerprint + beta merge
pipeline constructs all necessary headers from our controlled values.
Also: remove systemIncludesClaudeCodePrompt gate — OAuth accounts
now unconditionally rewrite system (even if client already sent a
Claude Code-style prompt), ensuring billing attribution block is
always present.
Before: isClaudeCodeRequest() checked whether the client looks like a
real Claude Code CLI (UA, system prompt, X-App header, metadata format).
If it looked like Claude Code, all mimicry was skipped — the assumption
being that a real CLI needs no help.
Problem: third-party tools like opencode partially impersonate Claude
Code (sending claude-cli UA + claude-code beta + CC system prompt) but
miss critical details (billing attribution block, tool-name obfuscation,
cache breakpoints, full beta set). Some users' opencode instances pass
the isClaudeCodeRequest check, causing sub2api to skip mimicry entirely,
while Anthropic still detects the request as third-party.
This explains why 'same opencode version, some users work, some don't'
— it depends on which opencode features/config trigger the validator.
Fix: OAuth accounts now unconditionally run the full mimicry pipeline,
matching Parrot's behavior (Parrot never checks client identity).
This is safe because our mimicry is strictly more complete than any
third-party client's partial impersonation.
Changed:
- /v1/messages path: remove isClaudeCode gate
- /v1/messages/count_tokens path: same
The previous commit only wired stripMessageCacheControl,
addMessageCacheBreakpoints, and tool-name obfuscation into
applyClaudeCodeOAuthMimicryToBody (used by /chat/completions and
/responses). The native /v1/messages path and count_tokens path
have their own independent mimicry code blocks and were missed.
Now all three entry points share the same D/E/F pipeline:
- /v1/messages (gateway_service.go forwardAnthropic)
- /v1/messages/count_tokens (gateway_service.go countTokens)
- OpenAI compat (applyClaudeCodeOAuthMimicryToBody)
Implements the remaining three parity items with Parrot cc_mimicry:
D) Tool-name obfuscation
- Dynamic mapping when tools.length > 5 (matches Parrot threshold).
Fake names follow {prefix}{name[:3]}{i:02d} (e.g. 'manage_bas00').
Go port of random.Random(hash(tuple(names))) uses fnv64a seed +
math/rand; byte-exact reproduction is impossible (Python hash vs
Go hash), but the two invariants that matter are preserved:
* same input tool_names yield identical mapping (cache hit)
* prefix pool is shuffled (names look distributed)
- Static prefix map (sessions_ -> cc_sess_, session_ -> cc_ses_)
applied as fallback, matching Parrot TOOL_NAME_REWRITES verbatim.
- Server tools (web_search_20250305, computer_*, etc.) are NOT
renamed; only type=='function' and type=='custom' tools are.
- tool_choice.name is rewritten in sync (only when type=='tool').
- Response side: bytes-level replace on every SSE chunk / JSON
body at 6 injection points (standard stream/non-stream,
passthrough stream/non-stream, chat_completions stream +
non-stream, responses stream + non-stream). Reverse mapping
applied longest-fake-name-first to prevent substring conflicts
(parity with Parrot _restore_tool_names_in_chunk).
- tool_choice is no longer unconditionally deleted in
normalizeClaudeOAuthRequestBody — Parrot passes it through.
E) tools[-1] cache_control breakpoint
- Injected as {type:ephemeral, ttl:<DefaultCacheControlTTL>} when
the last tool has no cache_control. Client-provided ttl is
passed through unchanged (repo-wide policy).
F) messages cache_control strategy
- stripMessageCacheControl removes every client-provided
messages[*].content[*].cache_control (multi-turn stability).
- addMessageCacheBreakpoints then injects two stable breakpoints:
(1) last message, and (2) second-to-last user turn when
messages.length >= 4.
- Combined with the system block breakpoint and tools[-1]
breakpoint, this gives exactly the 4 breakpoints Anthropic
allows per request.
Non-trivial implementation details to be aware of when rebasing:
* Two new files, no upstream collision:
gateway_tool_rewrite.go (D + E algorithms)
gateway_messages_cache.go (F strip + breakpoints)
* Two new feature calls bolted onto the tail of
applyClaudeCodeOAuthMimicryToBody in gateway_service.go — rebase
conflicts will be ~10 lines maximum.
* Response-side injection points all wrap their existing write with
reverseToolNamesIfPresent(c, ...), preserving original behavior
when no mapping is stored (static prefix rollback still runs).
* Non-stream chat/responses switched from c.JSON to
json.Marshal + c.Data so bytes-level replace is possible.
* Retry bodies (FilterThinkingBlocksForRetry,
FilterSignatureSensitiveBlocksForRetry, RectifyThinkingBudget)
only prune blocks — they preserve the already-obfuscated tool
names, so no extra mapping re-application is needed.
Manual QA: end-to-end scenario verified with 6 tools (above threshold)
and tool_choice.type=='tool'. Obfuscation + restore roundtrip shown
in test logs; then removed the temp test file.
Tests (16 new):
- buildDynamicToolMap stability + below-threshold guard
- sanitizeToolName precedence (dynamic > static)
- restoreToolNamesInBytes longest-first + static rollback
- applyToolNameRewriteToBody skips server tools + syncs tool_choice
- applyToolsLastCacheBreakpoint defaults to 5m + passes client ttl
- stripMessageCacheControl + addMessageCacheBreakpoints in the
1/4/string-content cases + second-to-last user turn selection
- buildToolNameRewriteFromBody ReverseOrdered is desc-by-fake-length
- fake name shape follows Parrot {prefix}{head3}{i:02d}
Three field-level alignments in normalizeClaudeOAuthRequestBody to
match real Claude Code CLI traffic byte-for-byte:
1. temperature: previously deleted unconditionally; now passes
through client value, defaults to 1 when absent (real CLI
always sends temperature, default 1).
2. max_tokens: defaults to 128000 when absent (real CLI default).
3. context_management: when thinking.type is enabled/adaptive
and the client did not provide context_management, inject
{"edits":[{"type":"clear_thinking_20251015","keep":"all"}]}
to mirror real CLI behavior.
tool_choice removal is unchanged (Claude Code OAuth credentials
do not allow client-supplied tool_choice).
Tests updated:
- gateway_body_order_test.go: temperature/max_tokens are now
expected in output; tool_choice still removed.
- gateway_prompt_test.go: system array is now 2 blocks
(billing + cc prompt), assertions adjusted.
- gateway_anthropic_apikey_passthrough_test.go: same 2-block
assertion.
Real Claude Code CLI always sends a 2-block system array:
[0] {"type":"text", "text":"x-anthropic-billing-header: cc_version=X.Y.Z.{fp}; cc_entrypoint=cli; cch=00000;"}
[1] {"type":"text", "text":"You are Claude Code...", "cache_control":{...}}
Before this commit, sub2api's mimicry path only produced block [1].
The missing billing block is one of the primary third-party detection
signals Anthropic uses for Claude-Code-scoped OAuth tokens.
New file gateway_billing_block.go ports the fingerprint algorithm
(byte-for-byte from Parrot cc_mimicry.py:compute_fingerprint):
pick chars at positions [4,7,20] of the first user text, then
`sha256(SALT + chars + cc_version)[:3]`.
- claude/constants.go: CLICurrentVersion = "2.1.92" (must match UA)
- gateway_billing_block.go: computeClaudeCodeFingerprint +
buildBillingAttributionBlockJSON + extractFirstUserText
- gateway_service.go: rewriteSystemForNonClaudeCode now emits both
blocks in order; cch=00000 is filled in later by
signBillingHeaderCCH in buildUpstreamRequest.
Downstream compat note: syncBillingHeaderVersion's regex
`cc_version=\d+\.\d+\.\d+` only matches the semver triple,
leaving the `.{fp}` suffix intact when rewriting in buildUpstreamRequest.
Real Claude CLI traffic sends cache_control as
`{"type":"ephemeral","ttl":"1h"}`. Our previous payload only
sent `{"type":"ephemeral"}`, which is a bytewise mismatch with
the official CLI and one more third-party detection signal.
Policy: client-provided ttl is always passed through unchanged.
Proxy-generated cache_control blocks default to 5m (vs Parrot's 1h)
to avoid burning the 1h cache budget on automatic breakpoints while
still aligning with the `ttl` field being present.
- claude/constants.go: DefaultCacheControlTTL = "5m"
- apicompat/types.go: new AnthropicCacheControl type with TTL field;
AnthropicTool gains optional CacheControl pointer so the mimicry
path can attach a cache breakpoint to tools[-1] later.
- service/gateway_service.go: anthropicCacheControlPayload gains TTL;
marshalAnthropicSystemTextBlock and rewriteSystemForNonClaudeCode
emit ttl=5m by default.
The previous commit added FullClaudeCodeMimicryBetas() but the two
call sites in buildUpstreamRequest still hardcoded the old 3-token
subset. Anthropic now checks the complete set of beta tokens to
decide if a request qualifies as Claude Code. Wire them up:
- /v1/messages mimic path: requiredBetas = FullClaudeCodeMimicryBetas()
- /v1/messages/count_tokens mimic path: same + BetaTokenCounting
Haiku models keep the 2-token exemption (BetaOAuth + InterleaveThinking).
Before: the OpenAI-compat forwarders only called injectClaudeCodePrompt,
which prepends the Claude Code banner but leaves the rest of the body
in its original non-Claude-Code shape. The codebase already admits this
is insufficient (see the comment on rewriteSystemForNonClaudeCode in
gateway_service.go: "仅前置追加 Claude Code 提示词无法通过检测").
Effect: OAuth accounts served through /v1/chat/completions or /v1/responses
were detected as third-party apps and bled plan quota with:
Third-party apps now draw from your extra usage, not your plan limits.
Fix:
- apicompat.AnthropicRequest: add Metadata json.RawMessage so metadata
survives the OpenAI->Anthropic->Marshal round trip; without it the
downstream rewrite has no user_id to work with.
- service: extract applyClaudeCodeOAuthMimicryToBody, a ParsedRequest-free
variant of the /v1/messages mimicry pipeline
(rewriteSystemForNonClaudeCode + normalizeClaudeOAuthRequestBody +
metadata.user_id injection) so the OpenAI-compat forwarders can reuse it.
- service: add buildOAuthMetadataUserIDFromBody + hashBodyForSessionSeed
for the same reason (no ParsedRequest at the call site).
- ForwardAsChatCompletions / ForwardAsResponses: replace the 3-line
prompt-prepend with the full mimicry pipeline.
- applyClaudeCodeMimicHeaders: set x-client-request-id per-request
(real Claude CLI always does); missing/duplicated values are one more
third-party fingerprint signal.
No change to the native /v1/messages path: it already called the full
pipeline, we only lift those helpers into a reusable function.
Tests:
- go build ./... passes
- go test ./internal/service/... ./internal/pkg/apicompat/... passes
- lsp_diagnostics clean on all touched files
- pre-existing failures in internal/config are unrelated (env-sensitive
tests that also fail on upstream main)
Add quota exceeded check to IsSchedulable() and refactor
shouldClearStickySession to delegate to IsSchedulable(), eliminating
duplicated logic and fixing missed overload/rate-limit/expired checks.
Frontend displays quota exceeded status independently via quota fields.
Subscription-mode billing was consuming quota at TotalCost (raw) instead of
ActualCost (TotalCost * RateMultiplier), so per-group rate multipliers —
including free subscriptions (multiplier = 0) — were silently ignored.
Switch the three subscription cost writes in buildUsageBillingCommand,
finalizePostUsageBilling, and the legacy postUsageBilling fallback to
ActualCost, and add a table-driven test covering 2x / 0.5x / free multipliers
plus a balance-mode regression check.
Consolidates 9 call sites of resolveUpstreamResponseReadLimit + readUpstreamResponseBodyLimited + ErrUpstreamResponseBodyTooLarge error handling into a single ReadUpstreamResponseBody function with TooLargeWriter callback for API-format-specific error responses (Anthropic, OpenAI, countTokens).
Priority was wrong:
- Before: custom rules → LiteLLM (when ApplyPricingToAccountStats) → nil
- After: custom rules → totalCost (when ApplyPricingToAccountStats) → LiteLLM → nil
When ApplyPricingToAccountStats is enabled, use the request's actual
client billing cost (before multiplier) as account_stats_cost, instead
of recalculating from LiteLLM per-token prices which produced incorrect
values for per-request billing mode.
LiteLLM model pricing is now the final fallback (priority 3), used only
when neither custom rules nor ApplyPricingToAccountStats apply.