Add configurable Responses context projection

.env.example

@@ -46,6 +46,13 @@ CODEBUDDY2API_MAX_CONCURRENT=64
 # CODEBUDDY2API_MAX_INFLIGHT_PER_ACCOUNT=0
 # Optional session/attempt tracing; legacy preserves existing upstream header behavior.
 # CODEBUDDY2API_REQUEST_CONTEXT_MODE=legacy
+# Responses projection is hot and WebUI-editable unless explicitly set here.
+# balanced rewrites only recognized harness blocks with stable summaries; other text stays verbatim.
+# CODEBUDDY2API_RESPONSES_PROJECTION_MODE=balanced
+# Per-item head/tail limit for generated assistant text, tool arguments and tool results;
+# zero disables per-item trimming. Global request/output gates still apply.
+# Non-zero values must be between 256 and 33554432 bytes.
+# CODEBUDDY2API_RESPONSES_PROJECTION_MAX_BYTES=40000
 # compatible preserves aggregation; realtime forwards all protocol output incrementally
 # and does not regenerate malformed tool arguments.
 # CODEBUDDY2API_STREAM_MODE=compatible

README.md

@@ -85,6 +85,7 @@ The template binds to localhost only. Configure HTTPS and restrict network acces
 - **No Docker?** After installing dependencies and the WebUI, run `uv run converter.py` or `python3 converter.py` without `.env`. First local startup saves a default key and displays it once — see [Local Python setup](docs/deployment.md#local-python-setup).
 - **Where is my data?** Everything lives in `auth/` (or `/data/auth` in Docker): credentials, settings and log databases — see [Data and backups](docs/webui.md#data-and-backups).
 - **Which image tag should I use?** `latest` follows stable releases, `edge` follows main, version tags pin one release — see [Published images](docs/deployment.md#use-published-images).
+- **Responses tool output/arguments compressed or need verbatim passthrough?** Set `responses_projection_mode` (default `balanced`, or `passthrough` to disable projection) and `responses_projection_max_bytes` (default `40000`, `0` disables per-item trimming). Balanced trimming keeps the head/tail and reports original bytes, estimated tokens and total lines; client addresses stay unchanged — see [Responses projection](docs/clients.md#responses-projection) and [details](docs/advanced.md#responses-projection).
 
 ## Disclaimer
 

README.zh-CN.md

@@ -85,6 +85,7 @@ docker compose up -d --no-build
 - **没有 Docker？** 准备依赖和 WebUI 后，直接 `uv run converter.py` 或 `python3 converter.py`，无需 `.env`；首次本地启动保存默认 key 并仅在终端显示一次——见[本地 Python 运行](docs/deployment.zh-CN.md#本地-python-运行)。
 - **数据在哪里？** 全部位于 `auth/`（Docker 中为 `/data/auth`）：凭证、设置与日志数据库——见[数据与备份](docs/webui.zh-CN.md#数据与备份)。
 - **镜像标签怎么选？** `latest` 跟随稳定版，`edge` 跟随 main，版本标签固定某一发行版——见[使用已发布镜像](docs/deployment.zh-CN.md#使用已发布镜像)。
+- **Responses 工具输出/参数被压缩或需要完全原样？** 设置 `responses_projection_mode`（默认 `balanced`，设 `passthrough` 可完全关闭投影）及 `responses_projection_max_bytes`（默认 `40000`，`0` 禁用单项裁剪）。balanced 裁剪保留头尾并显示原始 bytes、估算 tokens 与总行数，客户端地址不变——见 [Responses 投影](docs/clients.zh-CN.md#responses-投影)及[详细说明](docs/advanced.zh-CN.md#responses-投影)。
 
 ## 免责声明
 

app/adapters/responses_adapter.py

@@ -160,14 +160,17 @@ def _flush_assistant():
 
         # Merge calls into the preceding assistant message.
         if item_type == "function_call":
+            arguments = item.get("arguments", "{}")
+            if not isinstance(arguments, str):
+                raise ValueError("function_call.arguments must be a JSON string")
             if pending_assistant_content is None:
                 pending_assistant_content = ""
             pending_tool_calls.append({
                 "id": item.get("call_id", item.get("id", _rand_id("call_"))),
                 "type": "function",
                 "function": {
                     "name": item.get("name", ""),
-                    "arguments": item.get("arguments", "{}"),
+                    "arguments": arguments,
                 },
             })
             continue