Bug type

Behavior bug (incorrect output/state without crash)

Beta release blocker

No

Summary

Bug type

Regression (worked before, now fails)

Summary

After updating to OpenClaw 2026.4.26, /new on openai-codex/gpt-5.5
immediately fails with:

One of "input" or "previous_response_id" or 'prompt' or 'conversation_id' must be provided.

/reset is also affected.

Environment

• OpenClaw: 2026.4.26 (be8c246)
• OS: macOS (Mac mini)
• Node: v25.8.1 (fnm)
• Install method: npm global
• Channel: Discord
• Primary model: openai-codex/gpt-5.5
• Auth: OpenAI Codex OAuth (ChatGPT Plus)

Steps to reproduce

1. Configure main agent with openai-codex/gpt-5.5 as primary model
2. Bind to Discord
3. Send /new (or /reset) in Discord
4. Bot replies with the "must be provided" error

Logs

Right after the error, gateway logs contain:

This indicates a blank user message was injected into the session and
reached the model adapter, but the repair only happens after the API
rejects the request — there's no pre-flight guard.

Suspected root cause

The 2026.4.25 changelog includes:

Sessions: keep embedded runtime context out of the visible user prompt
by sending it as a hidden next-turn custom message

On a fresh session created by /new, this hidden custom message
serializes to an empty input array on the OpenAI Responses API path,
and previous_response_id is null (new session), so the request fails
strict validation.

Anthropic and OpenAI Chat Completions accept this shape; OpenAI
Responses API does not.

Related issues (same family — Responses API edge cases not handled)

• OpenAI GPT-5.2: reasoning item without required following item #1158 (reasoning item orphan)
• 400 error: reasoning item provided without required following item (OpenAI Responses API) #19582 (reasoning item without following item)
• Codex (openai-codex-responses) fails: store=false causes 404 on tool result submission #16829 (store=false → 404 on previous_response_id)
• [Bug]:reasoning_content in conversation history causes oMLX JSON parse error on subsequent turns #46637 (reasoning_content control characters)
• OpenAI Codex OAuth token not passed to pi-ai streamSimpleOpenAICodexResponses (regression in 2026.3.26) #56364 (OAuth token not passed to pi-ai codex adapter)
• [Bug]: openai-codex/gpt-5.4 returns HTML payload instead of model output (webchat/TUI hangs, messages queue) #67624 (codex returns HTML payload)

Suggested fix

Add a pre-flight guard in the openai-codex-responses adapter (or
in the /new handler): if input serializes to empty AND
previous_response_id is null, either skip the model call entirely
(the boot turn doesn't need to be sent) or inject a minimal
placeholder input.

Steps to reproduce

use /new or /reset in discord.

Expected behavior

agent message.

Actual behavior

replies: One of "input" or "previous_response_id"or 'prompt'or 'conversation_id' must be provided.

OpenClaw version

OpenClaw: 2026.4.26 (be8c246)

Operating system

macOS 26.4.1

Install method

No response

Model

openai codex chatgpt 5.5

Provider / routing chain

not sure

Additional provider/model setup details

No response

Logs, screenshots, and evidence

Impact and severity

No response

Additional information

No response