Responses

POST http://127.0.0.1:8080/v1/responses, the API Codex CLI uses. Same model, tool calling and limits as Chat Completions.

curl

curl http://127.0.0.1:8080/v1/responses -H 'content-type: application/json' -d '{
  "model": "pooled",
  "instructions": "Answer in one sentence.",
  "input": "What is a KV cache?"
}'

Python

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:8080/v1", api_key="pooled")
r = client.responses.create(model="pooled", input="What is a KV cache?")
print(r.output_text)

# Continue from the stored response
r2 = client.responses.create(model="pooled", previous_response_id=r.id, input="Shorter.")
print(r2.output_text)

Input

input is a string or a list of items:

• messages with role user, assistant, system or developer
• function_call and function_call_output
• custom_tool_call and custom_tool_call_output
• reasoning: when encrypted_content holds a pooled1. value from this bridge, the room gets back its exact reasoning
• item_reference: an item of a response this bridge stored

instructions is the system prompt. As with OpenAI, it is not carried over through previous_response_id.

Parameters

Parameter	Behavior
max_output_tokens	Default 16384. Above 65536 is capped. Always capped by the context left.
temperature, top_k	As in Chat Completions.
reasoning.effort	Absent, none or minimal: thinking off. low to max: on.
include: ["reasoning.encrypted_content"]	Returns reasoning items you can send back exactly.
tools, tool_choice, parallel_tool_calls, max_tool_calls	Function and custom tools. max_tool_calls is 1 to 128; the room writes at most 16 calls per answer either way. See Tool calling.
text.format	json_object or json_schema. See Structured output.
previous_response_id, store	See below.
top_p, metadata	Accepted. top_p is ignored.

Stored responses and previous_response_id

With store (default true), the bridge keeps each response in memory. previous_response_id then continues from it, so you send only the new input.

• Up to 256 responses or 64 MB, whichever comes first. The least recently used goes first.
• Each is dropped an hour after its last use.
• Memory only: restarting the bridge forgets them.
• An unknown id is a 400 with code previous_response_not_found.

Stored responses also back these endpoints:

GET    /v1/responses/{id}
DELETE /v1/responses/{id}
GET    /v1/responses/{id}/input_items

Output

Output items come in the order they were generated: reasoning, then the message, then calls. status is completed, or incomplete when the answer hit max_output_tokens or the context. A call cut short that way is an incomplete function_call item.

usage has input_tokens_details.cached_tokens (prompt tokens the room already held) and output_tokens_details.reasoning_tokens.

Streaming

A stream is the full Responses event sequence, each event with an increasing sequence_number: response.created, response.in_progress, the output item, content part and delta events, and finally response.completed, response.incomplete or response.failed.

While the request waits or the room reads a long prompt, the bridge repeats response.in_progress every 10 s: real events, since clients such as Codex reset their idle timer only on events.

Reasoning streams as a reasoning item with a reasoning_text content part: response.reasoning_text.delta events, then response.reasoning_text.done.

Skipped and refused

Skipped, with one logged warning per client:

• hosted tools such as web_search (Codex always sends it)
• items from tools that ran elsewhere

A 400:

• a hosted tool in tool_choice
• a malformed allowed_tools entry
• background, conversation, prompt
• logprobs
• unknown item types
• GET /v1/responses/{id}?stream=true

/v1/responses/input_tokens and /v1/responses/compact are 404.