zerogravity/docs/endpoint-gap-analysis.md

Endpoint Gap Analysis

Updated: 2026-02-15
Sources: OpenAI Chat Completions API, OpenAI Responses API, Gemini Thinking Mode, proxy source code
Method: Full source audit cross-referenced against context7 OpenAI API docs

---

What's Implemented

All Endpoints

• ✅ Sync + streaming modes
• ✅ Model selection + validation
• ✅ OAuth auth check
• ✅ Timeout control
• ✅ Tool definitions, tool choice, tool results (OpenAI → Gemini auto-conversion)
• ✅ MITM bypass path for custom tools
• ✅ Thinking/reasoning in both sync and streaming
• ✅ Generation params forwarded via MITM (temperature, top_p, top_k, max_output_tokens, stop_sequences, frequency_penalty, presence_penalty)
• ✅ reasoning_effort / thinkingLevel — forwarded as generationConfig.thinkingConfig.thinkingLevel
• ✅ response_format: {type: "json_object"} — injected as responseMimeType: "application/json"
• ✅ Google Search grounding — web_search: true (Completions), tools: [{type: "web_search_preview"}] (Responses), google_search: true (Gemini)
• ✅ /v1/search endpoint — dedicated web search via Google Search grounding, returns structured results + citations
• ✅ Image uploads — input_image / image_url with base64 data URIs, injected via MITM as inlineData
• ✅ Upstream error propagation — Google API errors (400, 429, 500) returned to client instantly instead of hanging

Reasoning Effort → Thinking Level Mapping

OpenAI reasoning_effort	Google thinkingLevel	Gemini 3 Pro	Gemini 3 Flash
"low"	"low"	✅	✅
"medium"	"medium"	❌	✅
"high"	"high"	✅ (default)	✅ (default)
—	"minimal"	❌	✅

Completions-Specific

• ✅ stream_options.include_usage — final chunk with usage before [DONE]
• ✅ completion_tokens_details.reasoning_tokens — thinking token count
• ✅ prompt_tokens_details.cached_tokens — cache read tokens
• ✅ temperature, top_p, max_tokens, max_completion_tokens, frequency_penalty, presence_penalty
• ✅ reasoning_effort
• ✅ stop — string or array, forwarded as generationConfig.stopSequences
• ✅ response_format: {type: "json_object"} — injects responseMimeType
• ✅ response_format: {type: "json_schema", json_schema: {...}} — injects responseMimeType + responseSchema via MITM
• ✅ n (multiple choices) — fires N parallel cascades, collects into choices[] (sync only, capped at 5)
• ✅ conversation — session ID for multi-turn cascade reuse (custom extension)
• ✅ reasoning_content — thinking text in assistant message
• ✅ system_fingerprint — fp_<version> in sync + all streaming chunks
• ✅ service_tier — "default" in sync + all streaming chunks
• ✅ logprobs: null — in every choice (sync + streaming)
• ✅ metadata — accepted in request, ignored
• ✅ finish_reason — correctly maps Google's MAX_TOKENS→"length", SAFETY→"content_filter", etc.
• ✅ Full messages[] history — all user, assistant, system, tool messages forwarded

Responses-Specific

• ✅ Full streaming event set (all response.* events including reasoning summary)
• ✅ temperature, top_p, max_output_tokens
• ✅ reasoning_effort — echoed from client request
• ✅ thinking_signature for multi-turn thinking chains
• ✅ instructions, metadata, user — echoed in response
• ✅ Usage with MITM-intercepted real tokens
• ✅ max_tool_calls — limits tool calls returned per response
• ✅ conversation — session reuse
• ✅ previous_response_id, store, parallel_tool_calls, truncation, text.format, tool_choice — echoed
• ✅ tools — echoed from client request (was previously always [])
• ✅ text.format — {format: {type: "json_schema", ...}} injects responseMimeType + responseSchema via MITM, echoed in response

Gemini-Specific

• ✅ Native tool format (no conversion needed)
• ✅ usageMetadata in sync and streaming responses
• ✅ temperature, topP, topK, maxOutputTokens, stopSequences
• ✅ thinkingLevel
• ✅ Session/conversation reuse
• ✅ Array/multipart input — strings, string arrays, {text: "..."} object arrays

---

Fixed Bugs

#	Bug	Fix
B1	Messages history dropped	extract_chat_input now calls build_conversation_with_tools with ALL messages — full multi-turn via messages[] works.
B2	finish_reason never "length"	google_to_openai_finish_reason() helper maps MAX_TOKENS→"length", SAFETY/RECITATION/etc→"content_filter". Applied to all paths.
B3	reasoning always null	build_response_object now echoes client's reasoning_effort from RequestParams.
B4	tool_choice always "auto"	Changed from &'static str to serde_json::Value. Echoes whatever the client sent.
B5	tools always []	Echoes the client's tools array in the response.
B7	temperature/top_p wrong	Already defaults to 1.0 via unwrap_or(1.0). Was a false positive — no fix needed.

Acceptable / Won't Fix

#	Bug	Status
B6	Usage::estimate fake tokens as fallback	Only triggers on timeout/error paths. Heuristic len/4 is reasonable for timeouts where output tokens = 0.

---

TODO — New Features

Trivial (all done ✅)

All trivial response shape fixes have been implemented.

Medium (schema injection via MITM) — all done ✅

All structured output features have been implemented.

Hard (new features)

#	Gap	API	Notes
7	parallel_tool_calls	Both	Accept param, echo in response. Can't enforce server-side.

Stretch (research needed)

#	Gap	API	Notes
12	Audio input	Both	Audio modalities not yet supported. Vision/images work via MITM.

---

Won't Implement

#	Gap	Reason
9	prediction (Predicted Output)	Inference-level speculative decoding optimization. No Gemini equivalent.
10	logprobs / top_logprobs	Gemini never exposes token-level log probabilities.