b3nw/zerogravity
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: forward generation params via MITM + add usageMetadata to Gemini

Browse Source 
- Add GenerationParams struct to MitmStore for temperature, top_p,
  top_k, max_output_tokens, stop_sequences, frequency/presence_penalty
- MITM modify_request injects params into request.generationConfig
- All 3 endpoints (Completions, Responses, Gemini) store client params
- Add usageMetadata to Gemini sync responses (promptTokenCount,
  candidatesTokenCount, totalTokenCount, thoughtsTokenCount)
- Add generation param fields to GeminiRequest (temperature, topP, etc.)
- Completions stream_options.include_usage emits final usage chunk
- Completions reasoning_tokens in completion_tokens_details
- Update endpoint gap analysis doc (all high-priority gaps resolved)
This commit is contained in:
Nikketryhard
2026-02-15 14:23:05 -06:00
parent 735c3e357d
commit b1bd57ab5e
9 changed files with 1216 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
KNOWN_ISSUES.md
View File

@@ -72,11 +72,27 @@ heuristic hint is absent, properly correlating usage to cascades.
**Status: SOLVED (2026-02-15)**
The MITM proxy now captures `thinking_text` from `StreamingAccumulator` into
`MitmStore` as SSE chunks arrive. The Responses API streaming handler reads
thinking deltas from MitmStore and emits `response.reasoning_summary_text.delta`
events in real-time. This works for both Google (`thought: true` parts) and
Anthropic (`thinking_delta`) formats.
Thinking text now streams progressively as delta events. The implementation:
1. **LS cascade steps** — `plannerResponse.thinking` (field 3) grows progressively
as the LS receives data. For Opus 4.6, thinking text builds up word-by-word
over ~1-2s. For Gemini Flash, thinking arrives in 1-2 larger chunks.
2. **Delta tracking** — `last_thinking_len` tracks the previously emitted length.
Each poll compares current thinking length and emits only the new characters
as `response.reasoning_summary_text.delta` events.
3. **Lifecycle** — Structure events (`output_item.added`, `summary_part.added`)
emit on first thinking appearance. `done` events emit when response text
first appears (indicating thinking phase completed).
**Verified with Opus 4.6:** (2026-02-15 13:22 UTC)
```
delta_len=24 "The user is asking about"
delta_len=61 " the Collatz conjecture..."
delta_len=5 " This"
delta_len=10 " is a pure"
... (11 progressive deltas over ~850ms)
```
---

464
docs/endpoint-gap-analysis.md Normal file
View File

@@ -0,0 +1,464 @@
# Endpoint Gap Analysis
> **Generated:** 2026-02-15 (updated)
> **Proxy Version:** 3.1.0
> **Scope:** All three API endpoints vs official OpenAI / Gemini specifications
---
## Table of Contents
- [Endpoint Overview](#endpoint-overview)
- [Feature Parity Matrix](#feature-parity-matrix)
- [Detailed Endpoint Analysis](#detailed-endpoint-analysis)
- [Responses API (`/v1/responses`)](#responses-api-v1responses)
- [Chat Completions API (`/v1/chat/completions`)](#chat-completions-api-v1chatcompletions)
- [Gemini API (`/v1/gemini`)](#gemini-api-v1gemini)
- [Priority Gaps](#priority-gaps)
- [Architecture Notes](#architecture-notes)
---
## Endpoint Overview
The proxy exposes three main API endpoints, each serving different client ecosystems:
| Endpoint | Protocol | Primary Clients | Spec Reference |
| --------------------------- | --------------------------- | ----------------------------------------------------- | ------------------------------------------------------------------------------------------------------------ |
| `POST /v1/responses` | OpenAI Responses API | Claude Code, Antigravity-native clients | [platform.openai.com/docs/api-reference/responses](https://platform.openai.com/docs/api-reference/responses) |
| `POST /v1/chat/completions` | OpenAI Chat Completions API | OpenCode, Vercel AI SDK, any OpenAI-compatible client | [platform.openai.com/docs/api-reference/chat](https://platform.openai.com/docs/api-reference/chat/create) |
| `POST /v1/gemini` | Custom Gemini-native API | Direct Gemini-format consumers | [ai.google.dev/api](https://ai.google.dev/api) (loosely based) |
All three endpoints share the same backend pipeline:
```
Client Request → Proxy Endpoint → LS (Language Server) → Google API
↓
MITM Proxy (captures real usage + injects generation params + tool calls)
```
---
## Feature Parity Matrix
### Core Features
| Feature | Responses | Completions | Gemini |
| -------------------- | :-------: | :---------: | :----: |
| Sync mode | ✅ | ✅ | ✅ |
| Streaming mode (SSE) | ✅ | ✅ | ✅ |
| Model selection | ✅ | ✅ | ✅ |
| Model validation | ✅ | ✅ | ✅ |
| Auth check (OAuth) | ✅ | ✅ | ✅ |
| Timeout control | ✅ | ✅ | ✅ |
### Generation Parameters (MITM-injected)
| Feature | Responses | Completions | Gemini |
| ------------------- | :-------: | :---------: | :----: |
| `temperature` | ✅ | ✅ | ✅ |
| `top_p` / `topP` | ✅ | ✅ | ✅ |
| `top_k` / `topK` | ❌ | ❌ | ✅ |
| `max_output_tokens` | ✅ | ✅ | ✅ |
| `stop_sequences` | ❌ | ❌ | ✅ |
| `frequency_penalty` | ❌ | ✅ | ❌ |
| `presence_penalty` | ❌ | ✅ | ❌ |
> **Note:** All generation parameters are forwarded to Google's API via MITM injection into `request.generationConfig`. They override the LS defaults.
### Thinking / Reasoning
| Feature | Responses | Completions | Gemini |
| ---------------------------------- | :-------------------------------: | :-------------------------------: | :---------------------: |
| Thinking — LS path (streaming) | ✅ `reasoning_summary_text.delta` | ✅ `reasoning_content` delta | ✅ `thought: true` part |
| Thinking — LS path (sync) | ✅ `reasoning` output item | ✅ `reasoning_content` in message | ✅ `thought: true` part |
| Thinking — Bypass path (streaming) | ✅ | ✅ | ✅ |
| Thinking — Bypass path (sync) | ✅ | ✅ | ✅ |
| Thinking signature (multi-turn) | ✅ `thinking_signature` field | ❌ Not applicable | ❌ Not applicable |
### Tool Calls
| Feature | Responses | Completions | Gemini |
| ---------------------------- | :-----------------------------: | :------------------------: | :-------------------------------------: |
| Tool definitions input | ✅ OpenAI format → Gemini | ✅ OpenAI format → Gemini | ✅ Native Gemini format |
| Tool choice control | ✅ `tool_choice` | ✅ `tool_choice` | ✅ `tool_config` |
| Tool call output (streaming) | ✅ `function_call` items | ✅ `tool_calls` in delta | ✅ `functionCall` parts |
| Tool call output (sync) | ✅ `function_call` items | ✅ `tool_calls` in message | ✅ `functionCall` parts |
| Tool result input | ✅ `function_call_output` items | ✅ `tool` role messages | ✅ `functionResponse` in `tool_results` |
| MITM bypass (custom tools) | ✅ | ✅ | ✅ |
| Stale state protection | ✅ | ✅ | ✅ |
### Session Management
| Feature | Responses | Completions | Gemini |
| ------------------------------------ | :---------------------: | :--------------: | :---------------------: |
| Session/conversation reuse | ✅ `conversation` field | ❌ Not supported | ✅ `conversation` field |
| Session listing (`GET /v1/sessions`) | ✅ Shared | ✅ Shared | ✅ Shared |
| Session deletion | ✅ Shared | ✅ Shared | ✅ Shared |
### Usage / Token Tracking
| Feature | Responses | Completions | Gemini |
| -------------------------------- | :---------------------------: | :-------------------------------: | :--------------------------: |
| Usage in sync response | ✅ MITM real tokens | ✅ MITM real tokens | ✅ `usageMetadata` |
| Usage in streaming (final chunk) | ❌ Not emitted | ✅ `stream_options.include_usage` | ❌ Not emitted |
| `reasoning_tokens` in usage | ✅ In `output_tokens_details` | ✅ In `completion_tokens_details` | ✅ `thoughtsTokenCount` |
| Cache tokens | ✅ `cached_tokens` | ✅ `cached_tokens` | ✅ `cachedContentTokenCount` |
---
## Detailed Endpoint Analysis
### Responses API (`/v1/responses`)
**Spec:** [OpenAI Responses API](https://platform.openai.com/docs/api-reference/responses)
#### Request Fields
| Field | Spec | Status | Implementation Details |
| ---------------------------- | ------------- | :----: | -------------------------------------------------------------------------------------------------------------------------------- |
| `model` | Required | ✅ | Mapped to internal model enum via `lookup_model()` |
| `input` | Required | ✅ | String or array. Array supports `message` items and `function_call_output` items |
| `instructions` | Optional | ✅ | Prepended to user text as system instructions |
| `stream` | Optional | ✅ | SSE stream with `response.*` events |
| `tools` | Optional | ✅ | OpenAI function format → auto-converted to Gemini `functionDeclarations` via `openai_tools_to_gemini()` |
| `tool_choice` | Optional | ✅ | `"auto"`, `"required"`, `"none"`, or `{"type":"function","function":{"name":"X"}}` → converted to Gemini `functionCallingConfig` |
| `store` | Optional | ✅ | Accepted, echoed in response. Not actually persisted. |
| `temperature` | Optional | ✅ | **Forwarded** to Google via MITM `generationConfig` injection. |
| `top_p` | Optional | ✅ | **Forwarded** to Google via MITM. |
| `max_output_tokens` | Optional | ✅ | **Forwarded** to Google via MITM. |
| `previous_response_id` | Optional | ✅ | Accepted, echoed. Not used for chaining (use `conversation` instead). |
| `metadata` | Optional | ✅ | Accepted, echoed back in response. |
| `user` | Optional | ✅ | Accepted, echoed. |
| `conversation` | **Extension** | ✅ | Proxy-specific: session ID for multi-turn cascade reuse. |
| `timeout` | **Extension** | ✅ | Proxy-specific: request timeout in seconds (default 120). |
| `reasoning.effort` | Optional | ❌ | Could map to model variant selection (e.g., `"high"` → Opus, `"low"` → Flash). |
| `reasoning.generate_summary` | Optional | ❌ | Not implemented. Could control thinking output inclusion. |
| `truncation` | Optional | ❌ | Not applicable — LS manages context window. |
| `parallel_tool_calls` | Optional | ✅ | Hardcoded `true` in response. |
#### Response Object
| Field | Spec | Status | Notes |
| ---------------------- | ------------- | :----: | ---------------------------------------------------------------- |
| `id` | Required | ✅ | `resp_` + UUID |
| `object` | Required | ✅ | Always `"response"` |
| `created_at` | Required | ✅ | Unix timestamp |
| `status` | Required | ✅ | `"completed"` or `"incomplete"` |
| `completed_at` | Required | ✅ | Unix timestamp or null |
| `error` | Required | ✅ | null on success |
| `incomplete_details` | Required | ✅ | null |
| `instructions` | Required | ✅ | Echoed from request |
| `max_output_tokens` | Required | ✅ | Echoed or null |
| `model` | Required | ✅ | Model name string |
| `output` | Required | ✅ | Array of `reasoning` and/or `message` items |
| `parallel_tool_calls` | Required | ✅ | `true` |
| `previous_response_id` | Required | ✅ | Echoed or null |
| `reasoning` | Required | ✅ | `{effort: null, summary: null}` |
| `store` | Required | ✅ | Echoed |
| `temperature` | Required | ✅ | Echoed (default 1.0) |
| `text` | Required | ✅ | `{format: {type: "text"}}` |
| `tool_choice` | Required | ✅ | `"auto"` |
| `tools` | Required | ✅ | Echoed or `[]` |
| `top_p` | Required | ✅ | Echoed (default 1.0) |
| `truncation` | Required | ✅ | `"disabled"` |
| `usage` | Required | ✅ | MITM-intercepted real tokens when available, estimated otherwise |
| `user` | Required | ✅ | Echoed or null |
| `metadata` | Required | ✅ | Echoed or `{}` |
| `thinking_signature` | **Extension** | ✅ | Proxy-specific: opaque blob for multi-turn thinking chain |
#### Streaming Events
| Event | Spec | Status | Notes |
| ---------------------------------------- | --------- | :----: | ------------------------------ |
| `response.created` | Required | ✅ | Initial response shell |
| `response.in_progress` | Required | ✅ | |
| `response.output_item.added` | Required | ✅ | For reasoning + message items |
| `response.content_part.added` | Required | ✅ | |
| `response.output_text.delta` | Required | ✅ | Progressive text deltas |
| `response.output_text.done` | Required | ✅ | |
| `response.content_part.done` | Required | ✅ | |
| `response.output_item.done` | Required | ✅ | |
| `response.completed` | Required | ✅ | Final event with full response |
| `response.reasoning_summary_text.delta` | Required | ✅ | Progressive thinking deltas |
| `response.reasoning_summary_text.done` | Required | ✅ | |
| `response.function_call_arguments.delta` | For tools | ✅ | Tool call argument streaming |
| `response.function_call_arguments.done` | For tools | ✅ | |
---
### Chat Completions API (`/v1/chat/completions`)
**Spec:** [OpenAI Chat Completions API](https://platform.openai.com/docs/api-reference/chat/create)
#### Request Fields
| Field | Spec | Status | Implementation Details |
| ------------------------- | ------------- | :----: | -------------------------------------------------------------------- |
| `model` | Required | ✅ | Mapped to internal model enum |
| `messages` | Required | ✅ | Supports `system`, `developer`, `user`, `assistant`, `tool` roles |
| `messages[].content` | Required | ✅ | String or array of `{type: "text", text: "..."}` objects |
| `messages[].tool_calls` | Optional | ✅ | For assistant messages with tool calls |
| `messages[].tool_call_id` | Optional | ✅ | For tool result messages |
| `stream` | Optional | ✅ | SSE with `chat.completion.chunk` events |
| `stream_options` | Optional | ✅ | `include_usage: true` emits final usage chunk before `[DONE]` |
| `tools` | Optional | ✅ | OpenAI function format → auto-converted to Gemini |
| `tool_choice` | Optional | ✅ | `"auto"`, `"none"`, `"required"`, or specific function |
| `timeout` | **Extension** | ✅ | Proxy-specific (default 120s) |
| `temperature` | Optional | ✅ | **Forwarded** to Google via MITM `generationConfig.temperature` |
| `top_p` | Optional | ✅ | **Forwarded** to Google via MITM `generationConfig.topP` |
| `max_tokens` | Optional | ✅ | **Forwarded** to Google via MITM `generationConfig.maxOutputTokens` |
| `max_completion_tokens` | Optional | ✅ | **Forwarded** (same as `max_tokens`, newer OpenAI param) |
| `frequency_penalty` | Optional | ✅ | **Forwarded** to Google via MITM `generationConfig.frequencyPenalty` |
| `presence_penalty` | Optional | ✅ | **Forwarded** to Google via MITM `generationConfig.presencePenalty` |
| `user` | Optional | ✅ | Accepted, not used |
| `n` | Optional | ❌ | N/A — single generation only |
| `logprobs` | Optional | ❌ | N/A |
| `top_logprobs` | Optional | ❌ | N/A |
| `logit_bias` | Optional | ❌ | N/A |
| `response_format` | Optional | ❌ | Could be useful for JSON mode |
| `seed` | Optional | ❌ | N/A |
| `stop` | Optional | ❌ | Could be forwarded as `stopSequences` |
#### Sync Response Object
| Field | Spec | Status | Notes |
| ------------------------------------------------------------ | ----------- | :----: | -------------------------------------------- |
| `id` | Required | ✅ | `chatcmpl-` + UUID |
| `object` | Required | ✅ | `"chat.completion"` |
| `created` | Required | ✅ | Unix timestamp |
| `model` | Required | ✅ | Model name |
| `choices[0].index` | Required | ✅ | `0` |
| `choices[0].message.role` | Required | ✅ | `"assistant"` |
| `choices[0].message.content` | Required | ✅ | Response text |
| `choices[0].message.reasoning_content` | Extension | ✅ | Thinking text (when model produces thinking) |
| `choices[0].message.tool_calls` | Conditional | ✅ | When model returns tool calls |
| `choices[0].message.refusal` | Optional | ❌ | Not implemented |
| `choices[0].message.annotations` | Optional | ❌ | Not implemented |
| `choices[0].logprobs` | Optional | ❌ | Not implemented |
| `choices[0].finish_reason` | Required | ✅ | `"stop"` or `"tool_calls"` |
| `usage.prompt_tokens` | Required | ✅ | MITM real or estimated |
| `usage.completion_tokens` | Required | ✅ | MITM real or estimated |
| `usage.total_tokens` | Required | ✅ | Sum |
| `usage.prompt_tokens_details.cached_tokens` | Optional | ✅ | MITM cache read tokens |
| `usage.completion_tokens_details.reasoning_tokens` | Optional | ✅ | MITM thinking token count |
| `usage.completion_tokens_details.accepted_prediction_tokens` | Optional | ❌ | N/A |
| `usage.completion_tokens_details.rejected_prediction_tokens` | Optional | ❌ | N/A |
| `system_fingerprint` | Deprecated | ❌ | Cosmetic, not needed |
| `service_tier` | Optional | ❌ | Cosmetic, not needed |
#### Streaming Chunk Object
| Field | Spec | Status | Notes |
| ------------------------------------ | --------------- | :----: | ----------------------------------------------------- |
| `id` | Required | ✅ | Same across all chunks |
| `object` | Required | ✅ | `"chat.completion.chunk"` |
| `created` | Required | ✅ | Same across all chunks |
| `model` | Required | ✅ | |
| `choices[0].index` | Required | ✅ | `0` |
| `choices[0].delta.role` | First chunk | ✅ | `"assistant"` in first chunk |
| `choices[0].delta.content` | Text chunks | ✅ | Progressive text deltas |
| `choices[0].delta.reasoning_content` | Thinking chunks | ✅ | Progressive thinking deltas |
| `choices[0].delta.tool_calls` | Tool chunks | ✅ | Tool call data |
| `choices[0].delta` | Final chunk | ✅ | Empty `{}` |
| `choices[0].finish_reason` | Final chunk | ✅ | `"stop"` or `"tool_calls"` |
| `choices[0].logprobs` | Optional | ❌ | Not implemented |
| `usage` (final chunk) | Optional | ✅ | Emitted when `stream_options.include_usage` is `true` |
| `data: [DONE]` | Required | ✅ | Stream termination signal |
---
### Gemini API (`/v1/gemini`)
**Spec:** Custom endpoint loosely based on [Gemini REST API](https://ai.google.dev/api/generate-content)
> **Note:** This is NOT a 1:1 Gemini API replica. It's a simplified proxy-native endpoint that uses Gemini's `functionDeclarations` / `functionCall` / `functionResponse` format directly, avoiding OpenAI ↔ Gemini format conversion overhead.
#### Request Fields
| Field | Spec | Status | Implementation Details |
| --------------------------------------- | -------- | :----: | --------------------------------------------------------------- |
| `model` | Required | ✅ | Mapped to internal model enum |
| `input` | Required | ✅ | String only (no array/multipart) |
| `tools` | Optional | ✅ | Native Gemini `[{functionDeclarations: [...]}]` format |
| `tool_config` | Optional | ✅ | Native Gemini `{functionCallingConfig: {mode: "AUTO"}}` |
| `tool_results` | Optional | ✅ | Array of `{functionResponse: {name, response}}` |
| `conversation` | Optional | ✅ | Session ID for cascade reuse |
| `stream` | Optional | ✅ | SSE streaming |
| `timeout` | Optional | ✅ | Default 120s |
| `temperature` | Optional | ✅ | **Forwarded** to Google via MITM `generationConfig.temperature` |
| `top_p` / `topP` | Optional | ✅ | **Forwarded** to Google via MITM `generationConfig.topP` |
| `top_k` / `topK` | Optional | ✅ | **Forwarded** to Google via MITM `generationConfig.topK` |
| `max_output_tokens` / `maxOutputTokens` | Optional | ✅ | **Forwarded** via MITM `generationConfig.maxOutputTokens` |
| `stop_sequences` / `stopSequences` | Optional | ✅ | **Forwarded** via MITM `generationConfig.stopSequences` |
#### Sync Response Object
| Field | Spec | Status | Notes |
| -------------------------------------------- | --------- | :----: | -------------------------------- |
| `candidates[0].content.parts` | Required | ✅ | Array of text/functionCall parts |
| `candidates[0].content.parts[].text` | Text | ✅ | Response text |
| `candidates[0].content.parts[].thought` | Extension | ✅ | `true` for thinking parts |
| `candidates[0].content.parts[].functionCall` | Tool call | ✅ | `{name, args}` |
| `candidates[0].content.role` | Required | ✅ | `"model"` |
| `candidates[0].finishReason` | Required | ✅ | `"STOP"` |
| `modelVersion` | Required | ✅ | Model name string |
| `usageMetadata` | Optional | ✅ | MITM-intercepted token counts |
`usageMetadata` fields:
| Field | Status | Notes |
| ------------------------- | :----: | ------------------------- |
| `promptTokenCount` | ✅ | Input tokens |
| `candidatesTokenCount` | ✅ | Output tokens |
| `totalTokenCount` | ✅ | Input + output |
| `thoughtsTokenCount` | ✅ | Thinking/reasoning tokens |
| `cachedContentTokenCount` | ✅ | Cache read tokens |
#### Streaming Format
Each SSE `data:` chunk is a complete Gemini-format JSON object with progressive `candidates[0].content.parts`:
```
data: {"candidates":[{"content":{"parts":[{"text":"thinking...","thought":true}],"role":"model"}}],"modelVersion":"opus-4.6"}
data: {"candidates":[{"content":{"parts":[{"text":"Hello!"}],"role":"model"}}],"modelVersion":"opus-4.6"}
data: {"candidates":[{"content":{"parts":[{"text":""}],"role":"model"},"finishReason":"STOP"}],"modelVersion":"opus-4.6"}
data: [DONE]
```
---
## Priority Gaps
### 🔴 High Priority — RESOLVED ✅
All high-priority gaps have been addressed:
1. ~~**Completions: `stream_options.include_usage`**~~ → ✅ Implemented
2. ~~**Completions: `completion_tokens_details.reasoning_tokens`**~~ → ✅ Implemented
3. ~~**Completions: Accept `temperature`, `top_p`, `max_tokens`**~~ → ✅ Forwarded via MITM
4. ~~**Gemini: `usageMetadata`**~~ → ✅ Implemented
### 🟡 Medium Priority
5. **Responses: `reasoning.effort`**
- **What:** Map reasoning effort levels (`"high"`, `"medium"`, `"low"`) to model variant selection
- **Why:** Could automatically select Opus vs Flash based on reasoning needs
- **Effort:** Medium — needs model selection logic changes
6. **Completions: Session/conversation support**
- **What:** Add session reuse similar to Responses and Gemini endpoints
- **Why:** Would allow multi-turn conversations via the completions API
- **Effort:** Medium — need a way to pass session ID (maybe via `user` field or custom header)
7. **Completions: `stop` sequences**
- **What:** Forward `stop` to Google as `stopSequences` in `generationConfig`
- **Why:** Some clients use stop sequences to control generation
- **Effort:** Trivial — just add to `CompletionRequest` and `GenerationParams`
8. **Completions: `response_format` (JSON mode)**
- **What:** Forward `response_format: {"type": "json_object"}` to Google's `responseMimeType`
- **Why:** Useful for structured output
- **Effort:** Low — inject `responseMimeType: "application/json"` in generationConfig
### 🟢 Low Priority
Cosmetic or not applicable to our architecture:
9. **`system_fingerprint`** — OpenAI-specific field, meaningless for our proxy
10. **`service_tier`** — OpenAI billing concept, not applicable
11. **`n` > 1** — Multiple completions per request; our backend only generates one
12. **`logprobs`** — Would require token-level access we don't have
13. **`seed`** — Deterministic sampling not controllable through our proxy
---
## Architecture Notes
### Generation Parameter Injection
Client-specified sampling parameters are forwarded to Google's API via the MITM request modification pipeline:
```
Client sends temperature=0.5 → API handler stores in MitmStore.generation_params
↓
LS sends request to Google API
↓
MITM intercepts request
↓
modify_request() reads generation_params
↓
Injects into request.generationConfig:
temperature, topP, topK, maxOutputTokens,
stopSequences, frequencyPenalty, presencePenalty
↓
Forwards modified request to Google
```
This approach overrides whatever defaults the LS sets, giving clients direct control over sampling parameters.
### Dual Path Architecture
All three endpoints share a dual-path architecture:
```
┌─────────────────┐
│ Has custom │
Request ────────────► │ tools? │
└────────┬────────┘
│
┌──── Yes ──┴── No ────┐
│ │
┌─────▼─────┐ ┌─────▼─────┐
│ MITM │ │ LS Steps │
│ Bypass │ │ Polling │
│ Path │ │ Path │
└─────┬─────┘ └─────┬─────┘
│ │
┌─────▼─────┐ ┌─────▼─────┐
│ Poll │ │ Poll │
│ MitmStore │ │ get_steps │
│ directly │ │ from LS │
└─────┬─────┘ └─────┬─────┘
│ │
└──────────┬────────────┘
│
┌─────▼─────┐
│ Response │
│ to client │
└───────────┘
```
- **Bypass Path:** When custom tools are present, the handler polls `MitmStore` directly for response text, thinking text, and function calls. The MITM proxy captures these from the Google API response before the LS processes them.
- **LS Path:** When no custom tools are present, the handler polls the LS's `get_steps` API for progressive response data (text, thinking, status).
### Stale State Protection
All bypass paths include protection against stale `response_complete` flags from previous requests:
```rust
if complete && text.is_empty() && thinking.is_none() {
warn!("stale response_complete detected — clearing");
state.mitm_store.clear_response_async().await;
continue; // or retry
}
```
This handles the race condition where a previous request's MITM handler calls `mark_response_complete()` after the new request has already called `clear_response_async()`.
### Tool Format Conversion
```
OpenAI tools ──► openai_tools_to_gemini() ──► Gemini functionDeclarations
│
MitmStore.set_tools()
│
MITM proxy injects into
outgoing LS request
```
The Gemini endpoint skips this conversion entirely — tools are stored in native Gemini format.

242
src/api/completions.rs
View File

@@ -10,7 +10,7 @@ use std::sync::Arc;
use tracing::{debug, info, warn};
use super::models::{lookup_model, DEFAULT_MODEL, MODELS};
use super::polling::{extract_response_text, is_response_done, poll_for_response};
use super::polling::{extract_response_text, extract_thinking_content, is_response_done, poll_for_response};
use super::types::*;
use super::util::{err_response, now_unix};
use super::AppState;
@@ -176,6 +176,28 @@ pub(crate) async fn handle_completions(
}
state.mitm_store.clear_active_function_call();
// Store generation parameters for MITM injection
{
use crate::mitm::store::GenerationParams;
let gp = GenerationParams {
temperature: body.temperature,
top_p: body.top_p,
top_k: None, // OpenAI doesn't have top_k
max_output_tokens: body.max_tokens.or(body.max_completion_tokens),
stop_sequences: None, // TODO: body.stop
frequency_penalty: body.frequency_penalty,
presence_penalty: body.presence_penalty,
};
// Only store if at least one param is set
if gp.temperature.is_some() || gp.top_p.is_some() || gp.max_output_tokens.is_some()
|| gp.frequency_penalty.is_some() || gp.presence_penalty.is_some()
{
state.mitm_store.set_generation_params(gp).await;
} else {
state.mitm_store.clear_generation_params().await;
}
}
let token = state.backend.oauth_token().await;
if token.is_empty() {
return err_response(
@@ -241,6 +263,8 @@ pub(crate) async fn handle_completions(
uuid::Uuid::new_v4().to_string().replace('-', "")
);
let include_usage = body.stream_options.as_ref().map_or(false, |o| o.include_usage);
if body.stream {
chat_completions_stream(
state,
@@ -248,6 +272,7 @@ pub(crate) async fn handle_completions(
model_name.to_string(),
cascade_id,
body.timeout,
include_usage,
)
.await
} else {
@@ -271,6 +296,7 @@ async fn chat_completions_stream(
model_name: String,
cascade_id: String,
timeout: u64,
include_usage: bool,
) -> axum::response::Response {
let stream = async_stream::stream! {
let start = std::time::Instant::now();
@@ -294,6 +320,7 @@ async fn chat_completions_stream(
})).unwrap_or_default()));
let mut keepalive_counter: u64 = 0;
let mut last_thinking_len: usize = 0;
while start.elapsed().as_secs() < timeout {
// ── Check for MITM-captured function calls FIRST ──
@@ -342,6 +369,27 @@ async fn chat_completions_stream(
"finish_reason": "tool_calls",
}],
})).unwrap_or_default()));
if include_usage {
let mitm = state.mitm_store.take_usage(&cascade_id).await
.or(state.mitm_store.take_usage("_latest").await);
let (pt, ct, crt, tt) = if let Some(ref u) = mitm {
(u.input_tokens, u.output_tokens, u.cache_read_input_tokens, u.thinking_output_tokens)
} else { (0, 0, 0, 0) };
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"id": completion_id,
"object": "chat.completion.chunk",
"created": now_unix(),
"model": model_name,
"choices": [],
"usage": {
"prompt_tokens": pt,
"completion_tokens": ct,
"total_tokens": pt + ct,
"prompt_tokens_details": { "cached_tokens": crt },
"completion_tokens_details": { "reasoning_tokens": tt },
},
})).unwrap_or_default()));
}
yield Ok(Event::default().data("[DONE]"));
return;
}
@@ -349,7 +397,37 @@ async fn chat_completions_stream(
// ── Check for MITM-captured response text (bypass LS) ──
if has_custom_tools {
if let Some(text) = state.mitm_store.peek_response_text().await {
let peek = state.mitm_store.peek_response_text().await;
let complete = state.mitm_store.is_response_complete();
let has_fc = state.mitm_store.has_active_function_call();
if keepalive_counter % 10 == 0 || peek.is_some() || complete || has_fc {
debug!(
"Completions bypass poll: peek={}, complete={}, has_fc={}, last_text_len={}",
peek.as_ref().map(|t| t.len()).unwrap_or(0),
complete, has_fc, last_text.len()
);
}
// Stream thinking text as reasoning_content deltas (MITM bypass)
if let Some(tc) = state.mitm_store.peek_thinking_text().await {
if tc.len() > last_thinking_len {
let delta = &tc[last_thinking_len..];
last_thinking_len = tc.len();
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"id": completion_id,
"object": "chat.completion.chunk",
"created": now_unix(),
"model": model_name,
"choices": [{
"index": 0,
"delta": {"reasoning_content": delta},
"finish_reason": serde_json::Value::Null,
}],
})).unwrap_or_default()));
}
}
if let Some(text) = peek {
if !text.is_empty() && text != last_text {
let delta = if text.len() > last_text.len() && text.starts_with(&*last_text) {
text[last_text.len()..].to_string()
@@ -387,12 +465,33 @@ async fn chat_completions_stream(
"finish_reason": "stop",
}],
})).unwrap_or_default()));
if include_usage {
let mitm = state.mitm_store.take_usage(&cascade_id).await
.or(state.mitm_store.take_usage("_latest").await);
let (pt, ct, crt, tt) = if let Some(ref u) = mitm {
(u.input_tokens, u.output_tokens, u.cache_read_input_tokens, u.thinking_output_tokens)
} else { (0, 0, 0, 0) };
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"id": completion_id,
"object": "chat.completion.chunk",
"created": now_unix(),
"model": model_name,
"choices": [],
"usage": {
"prompt_tokens": pt,
"completion_tokens": ct,
"total_tokens": pt + ct,
"prompt_tokens_details": { "cached_tokens": crt },
"completion_tokens_details": { "reasoning_tokens": tt },
},
})).unwrap_or_default()));
}
yield Ok(Event::default().data("[DONE]"));
return;
}
} else if state.mitm_store.is_response_complete() {
// Response complete but no text — might be a tool call we already handled
// or an empty response. Give it a moment then bail.
} else if complete {
// Response complete but no text — might be a tool call arriving shortly,
// stale state from a previous request, or an empty response.
tokio::time::sleep(tokio::time::Duration::from_millis(200)).await;
// Re-check function calls one more time
let final_check = state.mitm_store.take_any_function_calls().await;
@@ -437,10 +536,35 @@ async fn chat_completions_stream(
"finish_reason": "tool_calls",
}],
})).unwrap_or_default()));
if include_usage {
let mitm = state.mitm_store.take_usage(&cascade_id).await
.or(state.mitm_store.take_usage("_latest").await);
let (pt, ct, crt, tt) = if let Some(ref u) = mitm {
(u.input_tokens, u.output_tokens, u.cache_read_input_tokens, u.thinking_output_tokens)
} else { (0, 0, 0, 0) };
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"id": completion_id,
"object": "chat.completion.chunk",
"created": now_unix(),
"model": model_name,
"choices": [],
"usage": {
"prompt_tokens": pt,
"completion_tokens": ct,
"total_tokens": pt + ct,
"prompt_tokens_details": { "cached_tokens": crt },
"completion_tokens_details": { "reasoning_tokens": tt },
},
})).unwrap_or_default()));
}
yield Ok(Event::default().data("[DONE]"));
return;
}
}
// No text and no function calls but complete=true: stale state.
// Clear the flag so we wait for the real response from this request.
warn!("Completions: stale response_complete detected (no text, no FC) — clearing");
state.mitm_store.clear_response_async().await;
}
// When using bypass mode, skip LS step polling
@@ -457,6 +581,26 @@ async fn chat_completions_stream(
if let Ok((status, data)) = state.backend.get_steps(&cascade_id).await {
if status == 200 {
if let Some(steps) = data["steps"].as_array() {
// Stream thinking deltas (reasoning_content)
if let Some(tc) = extract_thinking_content(steps) {
if tc.len() > last_thinking_len {
let delta = &tc[last_thinking_len..];
last_thinking_len = tc.len();
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"id": completion_id,
"object": "chat.completion.chunk",
"created": now_unix(),
"model": model_name,
"choices": [{
"index": 0,
"delta": {"reasoning_content": delta},
"finish_reason": serde_json::Value::Null,
}],
})).unwrap_or_default()));
}
}
let text = extract_response_text(steps);
if !text.is_empty() && text != last_text {
@@ -496,6 +640,27 @@ async fn chat_completions_stream(
"finish_reason": "stop",
}],
})).unwrap_or_default()));
if include_usage {
let mitm = state.mitm_store.take_usage(&cascade_id).await
.or(state.mitm_store.take_usage("_latest").await);
let (pt, ct, crt, tt) = if let Some(ref u) = mitm {
(u.input_tokens, u.output_tokens, u.cache_read_input_tokens, u.thinking_output_tokens)
} else { (0, 0, 0, 0) };
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"id": completion_id,
"object": "chat.completion.chunk",
"created": now_unix(),
"model": model_name,
"choices": [],
"usage": {
"prompt_tokens": pt,
"completion_tokens": ct,
"total_tokens": pt + ct,
"prompt_tokens_details": { "cached_tokens": crt },
"completion_tokens_details": { "reasoning_tokens": tt },
},
})).unwrap_or_default()));
}
yield Ok(Event::default().data("[DONE]"));
return;
}
@@ -519,6 +684,27 @@ async fn chat_completions_stream(
"finish_reason": "stop",
}],
})).unwrap_or_default()));
if include_usage {
let mitm = state.mitm_store.take_usage(&cascade_id).await
.or(state.mitm_store.take_usage("_latest").await);
let (pt, ct, crt, tt) = if let Some(ref u) = mitm {
(u.input_tokens, u.output_tokens, u.cache_read_input_tokens, u.thinking_output_tokens)
} else { (0, 0, 0, 0) };
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"id": completion_id,
"object": "chat.completion.chunk",
"created": now_unix(),
"model": model_name,
"choices": [],
"usage": {
"prompt_tokens": pt,
"completion_tokens": ct,
"total_tokens": pt + ct,
"prompt_tokens_details": { "cached_tokens": crt },
"completion_tokens_details": { "reasoning_tokens": tt },
},
})).unwrap_or_default()));
}
yield Ok(Event::default().data("[DONE]"));
return;
}
@@ -553,6 +739,27 @@ async fn chat_completions_stream(
"finish_reason": "stop",
}],
})).unwrap_or_default()));
if include_usage {
let mitm = state.mitm_store.take_usage(&cascade_id).await
.or(state.mitm_store.take_usage("_latest").await);
let (pt, ct, crt, tt) = if let Some(ref u) = mitm {
(u.input_tokens, u.output_tokens, u.cache_read_input_tokens, u.thinking_output_tokens)
} else { (0, 0, 0, 0) };
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"id": completion_id,
"object": "chat.completion.chunk",
"created": now_unix(),
"model": model_name,
"choices": [],
"usage": {
"prompt_tokens": pt,
"completion_tokens": ct,
"total_tokens": pt + ct,
"prompt_tokens_details": { "cached_tokens": crt },
"completion_tokens_details": { "reasoning_tokens": tt },
},
})).unwrap_or_default()));
}
yield Ok(Event::default().data("[DONE]"));
};
@@ -582,14 +789,23 @@ async fn chat_completions_sync(
Some(u) => Some(u),
None => state.mitm_store.take_usage("_latest").await,
};
let (prompt_tokens, completion_tokens, cached_tokens) = if let Some(mitm_usage) = mitm {
(mitm_usage.input_tokens, mitm_usage.output_tokens, mitm_usage.cache_read_input_tokens)
let (prompt_tokens, completion_tokens, cached_tokens, thinking_tokens) = if let Some(mitm_usage) = mitm {
(mitm_usage.input_tokens, mitm_usage.output_tokens, mitm_usage.cache_read_input_tokens, mitm_usage.thinking_output_tokens)
} else if let Some(u) = &result.usage {
(u.input_tokens, u.output_tokens, 0)
(u.input_tokens, u.output_tokens, 0, 0)
} else {
(0, 0, 0)
(0, 0, 0, 0)
};
// Build message object, including reasoning_content if thinking is present
let mut message = serde_json::json!({
"role": "assistant",
"content": result.text,
});
if let Some(ref thinking) = result.thinking {
message["reasoning_content"] = serde_json::json!(thinking);
}
Json(serde_json::json!({
"id": completion_id,
"object": "chat.completion",
@@ -597,10 +813,7 @@ async fn chat_completions_sync(
"model": model_name,
"choices": [{
"index": 0,
"message": {
"role": "assistant",
"content": result.text,
},
"message": message,
"finish_reason": "stop",
}],
"usage": {
@@ -610,6 +823,9 @@ async fn chat_completions_sync(
"prompt_tokens_details": {
"cached_tokens": cached_tokens,
},
"completion_tokens_details": {
"reasoning_tokens": thinking_tokens,
},
},
}))
.into_response()

327
src/api/gemini.rs
View File

@@ -3,17 +3,19 @@
//! Accepts tools in Gemini `functionDeclarations` format directly,
//! returns `functionCall` in Gemini format directly.
//! No OpenAI ↔ Gemini format conversion.
//! Supports both sync (`stream: false`) and streaming (`stream: true`) modes.
use axum::{
extract::State,
http::StatusCode,
response::{IntoResponse, Json},
response::{sse::Event, IntoResponse, Json, Sse},
};
use rand::Rng;
use std::sync::Arc;
use tracing::info;
use tracing::{info, warn};
use super::models::{lookup_model, DEFAULT_MODEL, MODELS};
use super::polling::poll_for_response;
use super::polling::{extract_response_text, extract_thinking_content, is_response_done, poll_for_response};
use super::util::err_response;
use super::AppState;
use crate::mitm::store::PendingToolResult;
@@ -40,6 +42,21 @@ pub(crate) struct GeminiRequest {
/// Tool results in Gemini format: [{"functionResponse": {"name": "...", "response": {...}}}]
#[serde(default)]
pub tool_results: Option<Vec<serde_json::Value>>,
/// Sampling temperature.
#[serde(default)]
pub temperature: Option<f64>,
/// Nucleus sampling (topP).
#[serde(default, alias = "topP")]
pub top_p: Option<f64>,
/// Top-K sampling.
#[serde(default, alias = "topK")]
pub top_k: Option<u32>,
/// Max output tokens.
#[serde(default, alias = "maxOutputTokens")]
pub max_output_tokens: Option<u64>,
/// Stop sequences.
#[serde(default, alias = "stopSequences")]
pub stop_sequences: Option<Vec<String>>,
}
fn default_timeout() -> u64 {
@@ -54,6 +71,30 @@ fn extract_conversation_id(conv: &Option<serde_json::Value>) -> Option<String> {
}
}
/// Build Gemini-format usageMetadata from MITM store.
async fn build_usage_metadata(
store: &crate::mitm::store::MitmStore,
cascade_id: &str,
) -> serde_json::Value {
let usage = store.take_usage(cascade_id).await
.or(store.take_usage("_latest").await);
if let Some(usage) = usage {
serde_json::json!({
"promptTokenCount": usage.input_tokens,
"candidatesTokenCount": usage.output_tokens,
"totalTokenCount": usage.input_tokens + usage.output_tokens,
"thoughtsTokenCount": usage.thinking_output_tokens,
"cachedContentTokenCount": usage.cache_read_input_tokens,
})
} else {
serde_json::json!({
"promptTokenCount": 0,
"candidatesTokenCount": 0,
"totalTokenCount": 0,
})
}
}
pub(crate) async fn handle_gemini(
State(state): State<Arc<AppState>>,
Json(body): Json<GeminiRequest>,
@@ -124,6 +165,27 @@ pub(crate) async fn handle_gemini(
info!(count = results.len(), "Stored Gemini-native tool results for MITM injection");
}
// Store generation parameters for MITM injection
{
use crate::mitm::store::GenerationParams;
let gp = GenerationParams {
temperature: body.temperature,
top_p: body.top_p,
top_k: body.top_k,
max_output_tokens: body.max_output_tokens,
stop_sequences: body.stop_sequences.clone(),
frequency_penalty: None,
presence_penalty: None,
};
if gp.temperature.is_some() || gp.top_p.is_some() || gp.top_k.is_some()
|| gp.max_output_tokens.is_some() || gp.stop_sequences.is_some()
{
state.mitm_store.set_generation_params(gp).await;
} else {
state.mitm_store.clear_generation_params().await;
}
}
// Session/conversation management
let session_id_str = extract_conversation_id(&body.conversation);
let cascade_id = if let Some(ref sid) = session_id_str {
@@ -184,6 +246,24 @@ pub(crate) async fn handle_gemini(
}
}
// Dispatch to sync or stream
let model_name = model_name.to_string();
let timeout = body.timeout;
if body.stream {
gemini_stream(state, model_name, cascade_id, timeout).await
} else {
gemini_sync(state, model_name, cascade_id, timeout).await
}
}
// ─── Sync ────────────────────────────────────────────────────────────────────
async fn gemini_sync(
state: Arc<AppState>,
model_name: String,
cascade_id: String,
timeout: u64,
) -> axum::response::Response {
let has_custom_tools = state.mitm_store.get_tools().await.is_some();
// Clear stale response
@@ -192,7 +272,7 @@ pub(crate) async fn handle_gemini(
// ── MITM bypass: when tools active, poll MitmStore directly ──
if has_custom_tools {
let start = std::time::Instant::now();
while start.elapsed().as_secs() < body.timeout {
while start.elapsed().as_secs() < timeout {
// Check for function calls
let captured = state.mitm_store.take_any_function_calls().await;
if let Some(ref calls) = captured {
@@ -226,15 +306,32 @@ pub(crate) async fn handle_gemini(
// Check for completed text response
if state.mitm_store.is_response_complete() {
let text = state.mitm_store.take_response_text().await.unwrap_or_default();
let thinking = state.mitm_store.take_thinking_text().await;
// Guard against stale response_complete with no data
if text.is_empty() && thinking.is_none() {
warn!("Gemini sync bypass: stale response_complete — clearing");
state.mitm_store.clear_response_async().await;
tokio::time::sleep(tokio::time::Duration::from_millis(200)).await;
continue;
}
let mut parts: Vec<serde_json::Value> = Vec::new();
if let Some(ref t) = thinking {
parts.push(serde_json::json!({"text": t, "thought": true}));
}
parts.push(serde_json::json!({"text": text}));
return Json(serde_json::json!({
"candidates": [{
"content": {
"parts": [{"text": text}],
"parts": parts,
"role": "model",
},
"finishReason": "STOP",
}],
"modelVersion": model_name,
"usageMetadata": build_usage_metadata(&state.mitm_store, &cascade_id).await,
}))
.into_response();
}
@@ -253,8 +350,7 @@ pub(crate) async fn handle_gemini(
}
// ── Normal LS path (no custom tools) ──
// Poll for response
let poll_result = poll_for_response(&state, &cascade_id, body.timeout).await;
let poll_result = poll_for_response(&state, &cascade_id, timeout).await;
// Check for captured function calls — return in Gemini format
let captured_tool_calls = state.mitm_store.take_any_function_calls().await;
@@ -287,20 +383,233 @@ pub(crate) async fn handle_gemini(
"finishReason": "STOP",
}],
"modelVersion": model_name,
"usageMetadata": build_usage_metadata(&state.mitm_store, &cascade_id).await,
}))
.into_response();
}
// Normal text response
// Normal text response — include thinking if present
let mut parts: Vec<serde_json::Value> = Vec::new();
if let Some(ref thinking) = poll_result.thinking {
parts.push(serde_json::json!({"text": thinking, "thought": true}));
}
parts.push(serde_json::json!({"text": poll_result.text}));
Json(serde_json::json!({
"candidates": [{
"content": {
"parts": [{"text": poll_result.text}],
"parts": parts,
"role": "model",
},
"finishReason": "STOP",
}],
"modelVersion": model_name,
"usageMetadata": build_usage_metadata(&state.mitm_store, &cascade_id).await,
}))
.into_response()
}
// ─── Streaming ───────────────────────────────────────────────────────────────
async fn gemini_stream(
state: Arc<AppState>,
model_name: String,
cascade_id: String,
timeout: u64,
) -> axum::response::Response {
let stream = async_stream::stream! {
let start = std::time::Instant::now();
let mut last_text = String::new();
let mut last_thinking = String::new();
let has_custom_tools = state.mitm_store.get_tools().await.is_some();
// Clear stale response
state.mitm_store.clear_response_async().await;
while start.elapsed().as_secs() < timeout {
// ── Check for MITM-captured function calls FIRST ──
let captured = state.mitm_store.take_any_function_calls().await;
if let Some(ref calls) = captured {
if !calls.is_empty() {
let parts: Vec<serde_json::Value> = calls
.iter()
.map(|fc| {
serde_json::json!({
"functionCall": {
"name": fc.name,
"args": fc.args,
}
})
})
.collect();
yield Ok::<_, std::convert::Infallible>(Event::default().data(serde_json::to_string(&serde_json::json!({
"candidates": [{
"content": {
"parts": parts,
"role": "model",
},
"finishReason": "STOP",
}],
"modelVersion": model_name,
})).unwrap_or_default()));
yield Ok(Event::default().data("[DONE]"));
return;
}
}
// ── MITM bypass path (custom tools active) ──
if has_custom_tools {
// Stream thinking
if let Some(tc) = state.mitm_store.peek_thinking_text().await {
if tc.len() > last_thinking.len() {
let delta = tc[last_thinking.len()..].to_string();
last_thinking = tc;
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"candidates": [{
"content": {
"parts": [{"text": delta, "thought": true}],
"role": "model",
},
}],
"modelVersion": model_name,
})).unwrap_or_default()));
}
}
// Stream text
if let Some(text) = state.mitm_store.peek_response_text().await {
if !text.is_empty() && text.len() > last_text.len() {
let delta = &text[last_text.len()..];
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"candidates": [{
"content": {
"parts": [{"text": delta}],
"role": "model",
},
}],
"modelVersion": model_name,
})).unwrap_or_default()));
last_text = text;
}
}
// Check completion
let complete = state.mitm_store.is_response_complete();
if complete && !last_text.is_empty() {
// Final chunk with finishReason
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"candidates": [{
"content": {
"parts": [{"text": ""}],
"role": "model",
},
"finishReason": "STOP",
}],
"modelVersion": model_name,
})).unwrap_or_default()));
yield Ok(Event::default().data("[DONE]"));
return;
} else if complete && last_text.is_empty() && last_thinking.is_empty() {
// Stale state — clear and retry
warn!("Gemini stream bypass: stale response_complete — clearing");
state.mitm_store.clear_response_async().await;
}
let poll_ms: u64 = rand::thread_rng().gen_range(150..300);
tokio::time::sleep(tokio::time::Duration::from_millis(poll_ms)).await;
continue;
}
// ── Normal LS path (no custom tools) ──
if let Ok((status, data)) = state.backend.get_steps(&cascade_id).await {
if status == 200 {
if let Some(steps) = data["steps"].as_array() {
// Stream thinking
if let Some(tc) = extract_thinking_content(steps) {
if tc.len() > last_thinking.len() {
let delta = tc[last_thinking.len()..].to_string();
last_thinking = tc;
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"candidates": [{
"content": {
"parts": [{"text": delta, "thought": true}],
"role": "model",
},
}],
"modelVersion": model_name,
})).unwrap_or_default()));
}
}
// Stream text
let text = extract_response_text(steps);
if !text.is_empty() && text.len() > last_text.len() {
let delta = if text.starts_with(&*last_text) {
&text[last_text.len()..]
} else {
&text
};
if !delta.is_empty() {
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"candidates": [{
"content": {
"parts": [{"text": delta}],
"role": "model",
},
}],
"modelVersion": model_name,
})).unwrap_or_default()));
last_text = text.to_string();
}
}
// Done check
if is_response_done(steps) && !last_text.is_empty() {
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"candidates": [{
"content": {
"parts": [{"text": ""}],
"role": "model",
},
"finishReason": "STOP",
}],
"modelVersion": model_name,
})).unwrap_or_default()));
yield Ok(Event::default().data("[DONE]"));
return;
}
}
}
}
let poll_ms: u64 = rand::thread_rng().gen_range(200..350);
tokio::time::sleep(tokio::time::Duration::from_millis(poll_ms)).await;
}
// Timeout
yield Ok(Event::default().data(serde_json::to_string(&serde_json::json!({
"candidates": [{
"content": {
"parts": [{"text": if last_text.is_empty() { "[Timeout]" } else { "" }}],
"role": "model",
},
"finishReason": "STOP",
}],
"modelVersion": model_name,
})).unwrap_or_default()));
yield Ok(Event::default().data("[DONE]"));
};
Sse::new(stream)
.keep_alive(
axum::response::sse::KeepAlive::new()
.interval(std::time::Duration::from_secs(15))
.text(""),
)
.into_response()
}

70
src/api/responses.rs
View File

@@ -242,6 +242,25 @@ pub(crate) async fn handle_responses(
state.mitm_store.set_tool_config(gemini_config).await;
}
// Store generation parameters for MITM injection
{
use crate::mitm::store::GenerationParams;
let gp = GenerationParams {
temperature: body.temperature,
top_p: body.top_p,
top_k: None,
max_output_tokens: body.max_output_tokens,
stop_sequences: None,
frequency_penalty: None,
presence_penalty: None,
};
if gp.temperature.is_some() || gp.top_p.is_some() || gp.max_output_tokens.is_some() {
state.mitm_store.set_generation_params(gp).await;
} else {
state.mitm_store.clear_generation_params().await;
}
}
let response_id = format!(
"resp_{}",
uuid::Uuid::new_v4().to_string().replace('-', "")
@@ -1003,19 +1022,21 @@ async fn handle_responses_stream(
}
};
let mut thinking_started = false;
let mut thinking_done = false;
let mut last_thinking_len: usize = 0;
while start.elapsed().as_secs() < timeout {
if let Ok((status, data)) = state.backend.get_steps(&cascade_id).await {
if status == 200 {
if let Some(steps) = data["steps"].as_array() {
// Check for thinking content (appears before response text)
if !thinking_emitted {
// ── Phase 1: Stream thinking deltas progressively ──
if let Some(tc) = extract_thinking_content(steps) {
thinking_text = Some(tc.clone());
thinking_emitted = true;
if !thinking_started {
// First time we see thinking — emit structure events
thinking_started = true;
// Emit full reasoning event sequence at output_index 0
yield Ok(responses_sse_event(
"response.output_item.added",
serde_json::json!({
@@ -1040,6 +1061,14 @@ async fn handle_responses_stream(
"part": { "type": "summary_text", "text": "" },
}),
));
}
// Emit delta if thinking text has grown
if tc.len() > last_thinking_len {
let delta = &tc[last_thinking_len..];
last_thinking_len = tc.len();
thinking_text = Some(tc.clone());
yield Ok(responses_sse_event(
"response.reasoning_summary_text.delta",
serde_json::json!({
@@ -1048,9 +1077,22 @@ async fn handle_responses_stream(
"item_id": &reasoning_id,
"output_index": 0,
"summary_index": 0,
"delta": &tc,
"delta": delta,
}),
));
}
}
// ── Phase 2: Stream text deltas ──
let text = extract_response_text(steps);
let msg_output_index: u32 = if thinking_started { 1 } else { 0 };
if !text.is_empty() && text != last_text {
// Finalize thinking when response text first appears
if thinking_started && !thinking_done {
thinking_done = true;
let final_thinking = thinking_text.clone().unwrap_or_default();
yield Ok(responses_sse_event(
"response.reasoning_summary_text.done",
serde_json::json!({
@@ -1059,7 +1101,7 @@ async fn handle_responses_stream(
"item_id": &reasoning_id,
"output_index": 0,
"summary_index": 0,
"text": &tc,
"text": &final_thinking,
}),
));
yield Ok(responses_sse_event(
@@ -1070,7 +1112,7 @@ async fn handle_responses_stream(
"item_id": &reasoning_id,
"output_index": 0,
"summary_index": 0,
"part": { "type": "summary_text", "text": &tc },
"part": { "type": "summary_text", "text": &final_thinking },
}),
));
yield Ok(responses_sse_event(
@@ -1084,19 +1126,13 @@ async fn handle_responses_stream(
"type": "reasoning",
"summary": [{
"type": "summary_text",
"text": &tc,
"text": &final_thinking,
}],
},
}),
));
}
}
// ── Phase 2: Stream text deltas ──
let text = extract_response_text(steps);
let msg_output_index: u32 = if thinking_emitted { 1 } else { 0 };
if !text.is_empty() && text != last_text {
// Emit message output_item.added on first text
if !message_started {
message_started = true;
@@ -1153,7 +1189,7 @@ async fn handle_responses_stream(
if is_response_done(steps) && !last_text.is_empty() {
debug!("Response done, text length={}", last_text.len());
let mu = extract_model_usage(steps);
let msg_idx: u32 = if thinking_emitted { 1 } else { 0 };
let msg_idx: u32 = if thinking_started { 1 } else { 0 };
let (usage, mitm_thinking) = usage_from_poll(&state.mitm_store, &cascade_id, &mu, &params.user_text, &last_text).await;
let ts = extract_thinking_signature(steps);
// Use already-captured thinking, or MITM thinking, or LS thinking
@@ -1179,7 +1215,7 @@ async fn handle_responses_stream(
if run_status.contains("IDLE") && !last_text.is_empty() {
debug!("Trajectory IDLE, text length={}", last_text.len());
let mu = extract_model_usage(steps);
let msg_idx: u32 = if thinking_emitted { 1 } else { 0 };
let msg_idx: u32 = if thinking_started { 1 } else { 0 };
let (usage, mitm_thinking) = usage_from_poll(&state.mitm_store, &cascade_id, &mu, &params.user_text, &last_text).await;
let ts = extract_thinking_signature(steps);
let tc = thinking_text.clone()

32
src/api/types.rs
View File

@@ -40,6 +40,14 @@ pub(crate) struct ResponsesRequest {
pub tool_choice: Option<serde_json::Value>,
}
/// Stream options for Chat Completions (controls usage emission in final chunk).
#[derive(Deserialize, Default)]
pub(crate) struct StreamOptions {
/// When true, emit a final chunk with usage statistics before [DONE].
#[serde(default)]
pub include_usage: bool,
}
/// Chat Completions request (OpenAI-compatible).
#[derive(Deserialize)]
pub(crate) struct CompletionRequest {
@@ -53,6 +61,30 @@ pub(crate) struct CompletionRequest {
pub tools: Option<Vec<serde_json::Value>>,
/// Tool choice: "auto", "none", "required", or {"type":"function","function":{"name":"..."}}
pub tool_choice: Option<serde_json::Value>,
/// Stream options — controls whether usage is emitted in the final streaming chunk.
#[serde(default)]
pub stream_options: Option<StreamOptions>,
/// Sampling temperature — forwarded to Google via MITM generationConfig injection.
#[serde(default)]
pub temperature: Option<f64>,
/// Nucleus sampling — forwarded to Google via MITM.
#[serde(default)]
pub top_p: Option<f64>,
/// Max tokens — forwarded to Google as maxOutputTokens via MITM.
#[serde(default)]
pub max_tokens: Option<u64>,
/// Max completion tokens — forwarded to Google as maxOutputTokens via MITM.
#[serde(default)]
pub max_completion_tokens: Option<u64>,
/// User identifier — accepted, not used.
#[serde(default)]
pub user: Option<String>,
/// Frequency penalty — forwarded to Google via MITM.
#[serde(default)]
pub frequency_penalty: Option<f64>,
/// Presence penalty — forwarded to Google via MITM.
#[serde(default)]
pub presence_penalty: Option<f64>,
}
#[derive(Deserialize)]

58
src/mitm/modify.rs
View File

@@ -26,6 +26,8 @@ pub struct ToolContext {
pub pending_results: Vec<PendingToolResult>,
/// Last captured function calls for history rewriting.
pub last_calls: Vec<CapturedFunctionCall>,
/// Client-specified generation parameters (temperature, top_p, etc.).
pub generation_params: Option<super::store::GenerationParams>,
}
/// Modify a streamGenerateContent request body in-place.
@@ -349,6 +351,62 @@ pub fn modify_request(body: &[u8], tool_ctx: Option<&ToolContext>) -> Option<Vec
}
}
// ── 5. Inject client-specified generation parameters ──────────────────
// These override the LS defaults (which are typically absent or conservative).
// Google generationConfig fields: temperature, topP, topK, maxOutputTokens,
// stopSequences, frequencyPenalty, presencePenalty.
if let Some(ref ctx) = tool_ctx {
if let Some(ref gp) = ctx.generation_params {
// Find or create generationConfig (same path as above)
let gc = if let Some(req) = json.get_mut("request").and_then(|v| v.as_object_mut()) {
Some(req.entry("generationConfig")
.or_insert_with(|| serde_json::json!({})))
} else {
json.as_object_mut().map(|o| {
o.entry("generationConfig")
.or_insert_with(|| serde_json::json!({}))
})
};
if let Some(gc) = gc.and_then(|v| v.as_object_mut()) {
let mut injected: Vec<String> = Vec::new();
if let Some(t) = gp.temperature {
gc.insert("temperature".to_string(), serde_json::json!(t));
injected.push(format!("temperature={t}"));
}
if let Some(p) = gp.top_p {
gc.insert("topP".to_string(), serde_json::json!(p));
injected.push(format!("topP={p}"));
}
if let Some(k) = gp.top_k {
gc.insert("topK".to_string(), serde_json::json!(k));
injected.push(format!("topK={k}"));
}
if let Some(m) = gp.max_output_tokens {
gc.insert("maxOutputTokens".to_string(), serde_json::json!(m));
injected.push(format!("maxOutputTokens={m}"));
}
if let Some(ref seqs) = gp.stop_sequences {
gc.insert("stopSequences".to_string(), serde_json::json!(seqs));
injected.push(format!("stopSequences({})", seqs.len()));
}
if let Some(fp) = gp.frequency_penalty {
gc.insert("frequencyPenalty".to_string(), serde_json::json!(fp));
injected.push(format!("frequencyPenalty={fp}"));
}
if let Some(pp) = gp.presence_penalty {
gc.insert("presencePenalty".to_string(), serde_json::json!(pp));
injected.push(format!("presencePenalty={pp}"));
}
if !injected.is_empty() {
changes.push(format!("inject generationConfig: {}", injected.join(", ")));
}
}
}
}
if changes.is_empty() {
return None; // Nothing modified
}

4
src/mitm/proxy.rs
View File

@@ -561,13 +561,15 @@ async fn handle_http_over_tls(
let tool_config = store.get_tool_config().await;
let pending_results = store.take_tool_results().await;
let last_calls = store.get_last_function_calls().await;
let generation_params = store.get_generation_params().await;
let tool_ctx = if tools.is_some() || !pending_results.is_empty() {
let tool_ctx = if tools.is_some() || !pending_results.is_empty() || generation_params.is_some() {
Some(super::modify::ToolContext {
tools,
tool_config,
pending_results,
last_calls,
generation_params,
})
} else {
None

37
src/mitm/store.rs
View File

@@ -60,6 +60,21 @@ pub struct PendingToolResult {
pub result: serde_json::Value,
}
/// Client-specified generation parameters for MITM injection.
/// Set by API handlers, consumed by the MITM modify layer.
#[derive(Debug, Clone, Default)]
pub struct GenerationParams {
pub temperature: Option<f64>,
pub top_p: Option<f64>,
pub top_k: Option<u32>,
pub max_output_tokens: Option<u64>,
pub stop_sequences: Option<Vec<String>>,
/// Frequency penalty (OpenAI) — mapped to frequencyPenalty in Gemini.
pub frequency_penalty: Option<f64>,
/// Presence penalty (OpenAI) — mapped to presencePenalty in Gemini.
pub presence_penalty: Option<f64>,
}
/// Thread-safe store for intercepted data.
///
/// Keyed by a unique request ID that we can correlate with cascade operations.
@@ -102,6 +117,10 @@ pub struct MitmStore {
captured_thinking_text: Arc<RwLock<Option<String>>>,
/// Whether the captured response is complete (finishReason received).
response_complete: Arc<AtomicBool>,
// ── Generation parameters for MITM injection ─────────────────────────
/// Client-specified sampling parameters to inject into Google API requests.
generation_params: Arc<RwLock<Option<GenerationParams>>>,
}
/// Aggregate statistics across all intercepted traffic.
@@ -144,6 +163,7 @@ impl MitmStore {
captured_response_text: Arc::new(RwLock::new(None)),
captured_thinking_text: Arc::new(RwLock::new(None)),
response_complete: Arc::new(AtomicBool::new(false)),
generation_params: Arc::new(RwLock::new(None)),
}
}
@@ -433,4 +453,21 @@ impl MitmStore {
pub async fn clear_active_cascade(&self) {
*self.active_cascade_id.write().await = None;
}
// ── Generation parameters ────────────────────────────────────────────
/// Store client-specified generation parameters for MITM injection.
pub async fn set_generation_params(&self, params: GenerationParams) {
*self.generation_params.write().await = Some(params);
}
/// Read current generation parameters (non-consuming).
pub async fn get_generation_params(&self) -> Option<GenerationParams> {
self.generation_params.read().await.clone()
}
/// Clear generation parameters.
pub async fn clear_generation_params(&self) {
*self.generation_params.write().await = None;
}
}