test(translator): add unit tests for OpenAI to Claude requests and tool result handling
- Introduced tests for converting OpenAI requests to Claude with text, base64 images, and URL images in tool results.
- Refactored `convertClaudeToolResultContent` and related functionality to properly handle raw content with images and text.
- Updated conversion logic to streamline image handling for both base64 and URL formats.
The instructions restore logic was originally needed when the proxy
injected custom instructions (per-model system prompts) into requests.
Since ac802a46 removed the injection system, the proxy no longer
modifies instructions before forwarding. The upstream response's
instructions field now matches the client's original value, making
the restore a no-op.
Also removes unused sjson import.
Closesrouter-for-me/CLIProxyAPI#1868
Claude's Tool Search feature (advanced-tool-use-2025-11-20 beta) adds
defer_loading field to tool definitions. When proxying Claude requests
to Codex or Gemini, this unknown field causes 400 errors upstream.
Strip defer_loading (and cache_control where missing) in all three
Claude-to-upstream translation paths:
- codex/claude: defer_loading + cache_control
- gemini-cli/claude: defer_loading
- gemini/claude: defer_loading
Fixes#1725, Fixes#1375
Add support for Claude's "adaptive" and "auto" thinking modes using `output_config.effort`. Introduce support for new effort level "max" in adaptive thinking. Update thinking logic, validate model capabilities, and extend converters and handling to ensure compatibility with adaptive modes. Adjust static model data with supported levels and refine handling across translators and executors.
## Problem
When using Antigravity Claude models through CLIProxyAPI, the thinking
chain (reasoning content) does not display in the Amp client.
## Root Cause
The Amp client sends `thinking: {"type": "auto"}` in its requests,
but `ConvertClaudeRequestToAntigravity` only handled `"enabled"` and
`"adaptive"` types in its switch statement. The `"auto"` type was
silently ignored, resulting in no `thinkingConfig` being set in the
translated Gemini request. Without `thinkingConfig`, the Antigravity
API returns responses without any thinking content.
Additionally, the Antigravity API for Claude models does not support
`thinkingBudget: -1` (auto mode sentinel). It requires a concrete
positive budget value. The fix uses 128000 as the budget for "auto"
mode, which `ApplyThinking` will then normalize to stay within the
model's actual limits (e.g., capped to `maxOutputTokens - 1`).
## Changes
### internal/translator/antigravity/claude/antigravity_claude_request.go
1. **Add "auto" case** to the thinking type switch statement.
Sets `thinkingBudget: 128000` and `includeThoughts: true`.
The budget is subsequently normalized by `ApplyThinking` based
on model-specific limits.
2. **Add "auto" to hasThinking check** so that interleaved thinking
hints are injected for tool-use scenarios when Amp sends
`thinking.type="auto"`.
### internal/registry/model_definitions_static_data.go
3. **Add Thinking configuration** for `claude-sonnet-4-6`,
`claude-sonnet-4-5`, and `claude-opus-4-6` in
`GetAntigravityModelConfig()` -- these were previously missing,
causing `ApplyThinking` to skip thinking config entirely.
## Testing
- Deployed to Railway test instance (cpa-thinking-test)
- Verified via debug logging that:
- Amp sends `thinking: {"type": "auto"}`
- CPA now translates this to `thinkingConfig: {thinkingBudget: 128000, includeThoughts: true}`
- `ApplyThinking` normalizes the budget to model-specific limits
- Antigravity API receives the correct thinkingConfig
Amp-Thread-ID: https://ampcode.com/threads/T-019ca511-710d-776d-a07c-4b750f871a93
Co-authored-by: Amp <amp@ampcode.com>
Move base64 image data from Claude tool_result into functionResponse.parts
as inlineData instead of outer sibling parts, preventing context bloat.
Unify all inlineData field naming to camelCase mimeType across Claude,
OpenAI, and Gemini translators. Add comprehensive edge case tests and
Gemini-side regression test for functionResponse.parts preservation.
Non-spark codex models (gpt-5.3-codex, gpt-5.2-codex) stream function call
arguments via multiple delta events followed by a done event. The done handler
unconditionally emitted the full arguments, duplicating what deltas already
streamed. This produced invalid double JSON that Claude Code couldn't parse,
causing tool calls to fail with missing parameters and infinite retry loops.
Add HasReceivedArgumentsDelta flag to track whether delta events were received.
The done handler now only emits arguments when no deltas preceded it (spark
models), while delta-based streaming continues to work for non-spark models.
Some Codex models (e.g. gpt-5.3-codex-spark) send function call arguments
in a single "done" event without preceding "delta" events. The streaming
translator only handled "delta" events, causing tool call arguments to be
lost — resulting in empty tool inputs and infinite retry loops in clients
like Claude Code.
Emit the full arguments from the "done" event as a single input_json_delta
so downstream clients receive the complete tool input.
sjson.SetRaw with an empty string produces malformed JSON (e.g. "result":}).
This happens when a Claude tool_result block has no content field, causing
functionResponseResult.Raw to be "". Guard against this by falling back to
sjson.Set with an empty string only when .Raw is empty.
The OpenAI Chat Completions translator was silently dropping
response.function_call_arguments.delta and
response.function_call_arguments.done Codex SSE events, meaning
tool call arguments were never streamed incrementally to clients.
Add proper handling mirroring the proven Claude translator pattern:
- response.output_item.added: announce tool call (id, name, empty args)
- response.function_call_arguments.delta: stream argument chunks
- response.function_call_arguments.done: emit full args if no deltas
- response.output_item.done: defensive fallback for backward compat
State tracking via HasReceivedArgumentsDelta and HasToolCallAnnounced
ensures no duplicate argument emission and correct behavior for models
like codex-spark that skip delta events entirely.
- Added support for single and array-based `content` cases.
- Enhanced `system_instruction` structure population logic.
- Improved handling of user role assignment for string-based `content`.
