openclaw - 💡(How to fix) Fix [Bug]: Telegram /new can trigger empty Codex Responses request and expose provider error [1 comments, 2 participants]

Bug type

Regression / channel runtime error

Summary

On OpenClaw 2026.4.26, sending /new in a Telegram direct chat can reset/create the session but then trigger an empty Codex Responses API request. The raw provider error is delivered back to Telegram:

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

This looks like the /new slash command path starts a fresh session, then accidentally starts an assistant run without any user input or previous response/conversation id. The provider error is not wrapped before being sent to the chat.

Steps to reproduce

1. Run OpenClaw gateway with Telegram enabled.
2. Configure the default model to Codex OAuth, e.g. openai-codex/gpt-5.5.
3. In a Telegram direct chat, send /new as a standalone message.
4. Observe the bot reply with the raw OpenAI/Codex Responses API validation error.

Expected behavior

/new should only create/switch to a fresh session, or at most send a local acknowledgement.

It should not call the model with an empty request. If a model call would have no input, previous_response_id, prompt, or conversation_id, OpenClaw should stop locally and return a user-readable reset acknowledgement or internal error.

Actual behavior

The session appears to be reset/created, but an assistant turn fails before producing content. The raw provider error is sent to Telegram:

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

Local observation from the affected run:

stopReason: error
input: 0
error: 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 on Apple Silicon / Mac mini

Install method

npm global package, gateway running as LaunchAgent

Model

Default config:

agents.defaults.model.primary = openai-codex/gpt-5.5
agents.defaults.model.fallbacks = [deepseek/deepseek-v4-pro]

The Telegram session had an OpenAI Codex auth profile override:

authProfileOverride = openai-codex:<profile>

Provider / routing chain

Telegram direct chat -> main agent -> OpenAI Codex OAuth / Responses API

Config file / key location

~/.openclaw/openclaw.json

Relevant model config shape:

{
  "agents": {
    "defaults": {
      "model": {
        "primary": "openai-codex/gpt-5.5",
        "fallbacks": ["deepseek/deepseek-v4-pro"]
      },
      "models": {
        "openai-codex/gpt-5.5": {
          "params": {
            "fastMode": true
          }
        }
      }
    }
  }
}

Logs, screenshots, and evidence

The error text was delivered directly as the Telegram bot reply:

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

The affected Telegram direct session did get a fresh session id after /new, so the reset path partially worked. The failure appears to be the follow-up assistant run request assembly.

Impact and severity

Medium/high for Telegram users using Codex OAuth models:

• /new is documented as the recovery path for polluted/stale sessions.
• Instead, the recovery path can surface a provider validation error to the end user.
• The message looks like a user mistake even though it is a runtime request assembly problem.

Possible fix direction

• Ensure /new / /reset slash handlers do not start a model run unless there is real user input.
• Add a guard before Codex Responses calls: if no input, previous_response_id, prompt, or conversation_id would be sent, do not call the provider.
• Wrap provider validation errors before delivering to chat channels.

Related issues

Possibly related but not the same failure mode:

• #49517
• #69599
• #47504