Provider API /v1/responses rejects Codex's tool_search_call / tool_search_output input items with 400 "Invalid input"

[fenton-xue]

Summary

POST /provider/v1/responses returns HTTP 400 {"error":{"message":"Invalid input","type":"invalid_request_error","param":"input"}} for any request whose input array contains a tool_search_call or tool_search_output item.

Codex emits these items natively on the Responses wire (wire_api = "responses"), so every turn in which the model performs a tool search is killed before the model runs, and the task ends as a system error. On deepseek/deepseek-v4.1-flash this makes every MCP / plugin tool unusable: Codex defers MCP tools (js, js_reset, ...) behind a tool search, so the first time the agent reaches for Chrome or Computer Use the conversation history gains a tool_search_call/tool_search_output pair and the next request 400s.

Expected Behavior

The endpoint should accept tool_search_call / tool_search_output input items, or at minimum accept the request and ignore them. They are client-executed history items, the same class as function_call / function_call_output, which the endpoint already accepts today.

Actual Behavior

The whole request is rejected with 400 before any model output. Body is 84 bytes:

{"error":{"message":"Invalid input","type":"invalid_request_error","param":"input"}}

Steps to reproduce the issue

Baseline (200 OK):

curl -sS https://api.commandcode.ai/provider/v1/responses \
  -H "Authorization: Bearer $CMDCODE_KEY" \
  -H 'content-type: application/json' \
  -d '{
    "model": "deepseek/deepseek-v4.1-flash",
    "stream": false,
    "max_output_tokens": 16,
    "input": [{"role":"user","content":[{"type":"input_text","text":"hi"}]}]
  }'

Now append one tool_search_call item to input:

"input": [
  {"role":"user","content":[{"type":"input_text","text":"hi"}]},
  {"type":"tool_search_call","call_id":"call_x1","status":"completed","execution":"client","arguments":{"query":"cua"}}
]

Result: HTTP 400, the 84-byte body above.

Same with tool_search_output:

{"type":"tool_search_output","call_id":"call_x1","status":"completed","execution":"client","tools":[]}

Result: HTTP 400, same body.

Matrix measured against https://api.commandcode.ai/provider/v1/responses with model deepseek/deepseek-v4.1-flash:

input contents	result
plain user message	200
+ tool_search_call (full fields)	400
+ tool_search_call, minimal fields (type/call_id/arguments), no tools array on the request	400
+ tool_search_call, arguments as a JSON string instead of an object	400
+ tool_search_output, tools: []	400
+ tool_search_output, one real function tool in tools	400
+ function_call and function_call_output	200

The trigger is the item type itself, not a missing or extra field.

Real-world impact (Codex Desktop 0.155.0-alpha.9.2, macOS, zsh)

Turn 01a0bcca-9210-7650-bc44-862ddf3a9baf:

1. The user asks the agent to drive Chrome. The agent's js / js_reset calls start coming back as unsupported call: js (the mcp__cua_repl namespace is missing from the emitted calls).
2. The model then issues one tool search (arguments: {"query": "cua_repl javascript execution browser control"}) and Codex writes a tool_search_call + tool_search_output pair into the conversation history.
3. The next request 400s and the turn ends as:

{"type":"task_complete","error":{"message":"{\"error\":{\"message\":\"Invalid input\",\"type\":\"invalid_request_error\",\"param\":\"input\"}}","codex_error_info":"other"}}

A parallel thread on the same provider that never triggered a tool search ran to completion, matching the repro above.

Command Code Version

Provider API (gateway), server-side. Client: Codex Desktop 0.155.0-alpha.9.2, wire_api = "responses", base_url = https://api.commandcode.ai/provider/v1, model deepseek/deepseek-v4.1-flash.

Operating System

macOS

Terminal/IDE

Codex Desktop

Shell

zsh

Session file (optional)

No response

Fix prompt (optional)

POST /provider/v1/responses in the Provider API gateway rejects any request whose input array contains a tool_search_call or tool_search_output item, returning 400 {"error":{"message":"Invalid input","param":"input"}}. The endpoint already accepts function_call / function_call_output items, so extend the same history-item handling to tool_search_call / tool_search_output: parse them as client-executed history items and either carry them through the upstream translation or drop them, instead of failing request validation.

Reproduce: POST a request with a plain user message -> 200; append one tool_search_call item -> 400. Exact bodies are in the issue.

Check: both item types return a normal completion, and a Codex turn that performs a tool search no longer dies with param: "input".

Additional context

• Related and already closed: Provider API /v1/responses rejects Codex's non-function tool types (tools.N.type) and breaks the endpoint for Codex #885 (tools.N.type on /v1/responses), Provider API rejects tool_search function calls from Codex clients with 400 "Missing required parameter: 'input[N].arguments'" #745 (tool_search invocation on /chat/completions), Provider API: add OpenAI Responses API support (POST /provider/v1/responses) #832 (Responses support). This is a follow-up on the same endpoint: the tools array is accepted now, but the input history items still are not.
• tool_search_output.tools[] can itself contain {"type":"namespace","name":"mcp__node_repl", ...} entries with nested function tools, so whatever translation is added may need to tolerate nested namespace entries as well.
• Happy to dogfood a preview build and return full request/response traces.