Follow-up to #710 (AISIX-Cloud#953).
#710 fixes the 400 by dropping context_management (and other untranslated Anthropic-only fields) on the /v1/messages cross-provider path. That restores compatibility but loses the context-editing semantics the client asked for.
LiteLLM (HEAD 88e03e5) does more:
- On its OpenAI/Azure path it routes
/v1/messages through the Responses API and translates context_management into OpenAI's native compaction param ({"edits":[{"type":"compact_20260112","trigger":{"type":"input_tokens","value":N}}]} → [{"type":"compaction","compact_threshold":N}]); non-compact edits are skipped.
- On its chat-completions path it polyfills context management inside the gateway (
experimental_pass_through/context_management/): applies clear_tool_uses/compact edits to the message history before dispatch (compact calls a summarization model), so the upstream never sees the field.
Options for us, in increasing effort:
clear_tool_uses polyfill only (pure message-history rewrite, no extra model call)- full polyfill incl.
compact (needs a summarization call + config for which model to use)
- Responses-API-style translation if/when we dispatch cross-provider traffic to
/v1/responses-capable upstreams
Until then, dropping keeps Claude Code working against GPT models (context editing degrades gracefully — the client just never sees its context shrink).
Follow-up to #710 (AISIX-Cloud#953).
#710 fixes the 400 by dropping
context_management(and other untranslated Anthropic-only fields) on the/v1/messagescross-provider path. That restores compatibility but loses the context-editing semantics the client asked for.LiteLLM (HEAD
88e03e5) does more:/v1/messagesthrough the Responses API and translatescontext_managementinto OpenAI's native compaction param ({"edits":[{"type":"compact_20260112","trigger":{"type":"input_tokens","value":N}}]}→[{"type":"compaction","compact_threshold":N}]); non-compact edits are skipped.experimental_pass_through/context_management/): appliesclear_tool_uses/compactedits to the message history before dispatch (compact calls a summarization model), so the upstream never sees the field.Options for us, in increasing effort:
clear_tool_usespolyfill only (pure message-history rewrite, no extra model call)compact(needs a summarization call + config for which model to use)/v1/responses-capable upstreamsUntil then, dropping keeps Claude Code working against GPT models (context editing degrades gracefully — the client just never sees its context shrink).