Resolve const-block conflict in openai_gateway_grok_cache.go:
keep #4590's client tool cache constants, drop grokFreeRolling24hTokenLimit
(moved to pkg/xai as IsGrokFreeRolling24hTokenLimit with legacy 2M support).
Round-trip Grok encrypted reasoning via Anthropic thinking.signature and
seed cache identity from Claude Code session IDs, so multi-turn /v1/messages
traffic can hit xAI prompt cache more like CLIProxyAPI. Also strip invalid
encrypted_content and retry on the Messages path when decrypt fails after
account switches.
Use *string for AnthropicResponse.StopReason so streaming message_start
serializes as JSON null instead of "", matching official Anthropic wire
format and avoiding strict client turn-finalization failures.
The Anthropic→Responses streaming converter omits two things the OpenAI
Responses wire format requires, which breaks clients that reconstruct the
response from the event stream (rather than just reading deltas).
1. response.content_part.added is never emitted.
A message item is opened with content: [], and the OpenAI SDK's
accumulating stream helper (client.responses.stream) only appends a
content part when it sees content_part.added. Without it, the next
output_text.delta indexes output.content[content_index] and raises:
File "openai/lib/streaming/responses/_responses.py", line 352,
in accumulate_event
content = output.content[event.content_index]
IndexError: list index out of range
Raw iteration (responses.create(stream=True)) does not accumulate and is
unaffected, which is why this went unnoticed.
2. response.completed carries Output: []ResponsesOutput{}.
get_final_response() and tracing integrations parse the terminal event's
response directly, so callers see an empty output_text even though the
deltas streamed correctly. This one is invisible when only watching the
stream render.
Also carries the full text on output_text.done / content_part.done (deltas
carry increments only, done events carry the whole part) and fills in
content/arguments/summary on output_item.done, which had the same
empty-payload issue.
Reproduced against a live Anthropic-platform group with the openai Python
SDK 2.46.0; Arize Phoenix's playground hits the same path. Verified before
(IndexError) and after (full text via get_final_response()).
Adds regression tests covering event ordering, done-event payloads, and the
terminal event's output for both text and tool calls.
Behind a reverse proxy (e.g. nginx with X-Real-IP), admin audit logs and
session IP/UA binding always recorded 127.0.0.1 because they hardcoded
the gin trusted_proxies chain, while API key IP restriction already
honored the "trust forwarded client IP" system setting.
- add ip.GetSecurityClientIP(c, trustForwarded) as the single source of
truth for security-sensitive client IP selection; API key auth
middlewares (main + google) refactored onto it with zero behavior change
- SessionBindingContext(cfg) now resolves the client IP via the same
toggle and injects it into the request context; token issuance,
binding enforcement and its mismatch audit record all read the
injected value, so issue/verify can never diverge
- audit log middleware and audit-log clear trace record the same
security client IP (middleware.SecurityClientIP), falling back to the
trusted proxy chain when the injection is absent
- settings UI hint (zh/en) documents the broadened toggle scope and the
one-time re-login after toggling while session binding is enabled
With the toggle off (default) behavior is byte-for-byte unchanged.
Co-Authored-By: Claude <noreply@anthropic.com>
Adversarial re-verification found one more divergence from the
double-conversion chain: an upstream 200 with empty choices (or a nil
response) left stop_reason as an empty string, while the old chain
reports end_turn. Derive the fallback from the content blocks — the
guard never fires when choices exist, since every finish_reason maps to
a non-empty stop_reason.
Also strengthen the equivalence tests: compare tool_use Content[].Input
and tool_call Function.Arguments across bridges, and add an
empty-choices stop_reason parity case.
Adversarial review of the direct bridge found divergences from the
double-conversion chain it replaces; all are now aligned and covered by
tests that fail against the previous implementation:
- flush argument fragments buffered before a deferred tool announcement,
and announce name-less tools at finalize, so no tool arguments are lost
when upstreams stream arguments before the name
- fold text-only user array content into a single string; parts form only
when an image requires it (strict chat upstreams reject array content)
- drop tool_choice pointing at undeclared tools and unknown choice types
- treat cache_write_tokens/cache_creation_tokens as alternate spellings
(prefer write), not additive
- generate a response id when the upstream omits one
- derive stop_reason from blocks for content_filter/unknown finish reasons
- emit input_json_delta "{}" when a tool block closes without argument
deltas
Three tests covering the CC→Responses→Anthropic finalize path:
- TestStreamingParallelToolUseNoGhostDelta: full end-to-end with two
parallel tool_calls, asserts every content_block_delta targets a
started block
- TestStreamingParallelToolUseSecondToolPackedArgsDone: focused test
where tool 1 streams deltas and tool 2 has packed .done args
- TestStreamingThreeParallelToolsAllPackedDone: three tools all with
packed .done, the most extreme case
All three fail on unfixed code (ghost deltas on wrong indices) and
pass after the fix.
Skip the Responses API intermediate representation on the /v1/messages
force-chat path. Previously every streaming token ran through two state
machines (CC→Responses→Anthropic); now it runs through one (CC→Anthropic).
New file backend/internal/pkg/apicompat/chatcompletions_anthropic_bridge.go:
- AnthropicToChatCompletionsRequest: request-side direct conversion
- ChatCompletionsResponseToAnthropic: non-stream response direct conversion
- ChatCompletionsChunkToAnthropicEvents + Finalize: single streaming state machine
openai_gateway_messages_chat_fallback.go rewired to use the direct bridge.
Existing 7 ForceChatCompletions end-to-end tests pass unchanged; 22 new
unit tests added including equivalence tests vs the double-conversion path.
resToAnthHandleFuncArgsDone used state.ContentBlockIndex directly
instead of looking up from OutputIndexToBlockIdx like
resToAnthHandleFuncArgsDelta does. When multiple tool_calls arrive
in parallel and arguments come as a packed .done (no prior delta),
the second+ tool would emit content_block_delta on an index that
was never content_block_start'ed, causing Claude Code to report
"Content block not found".
Fix: resolve block index from OutputIndexToBlockIdx, and skip the
delta if the block is already closed or the index doesn't match
the current open block.
Closes#4193
eyre