#Problem in Telegram bot

1 messages · Page 1 of 1 (latest)

worn kelp
#

400 Item 'rs_00bfa83f6c825873006992f450659881a3820c7f9993447764' of type 'reasoning' was provided without its required following item.

teal tinselBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

delicate orioleBOT
#

That error is coming from OpenAI’s Responses API being strict about the transcript format: it’s seeing a standalone reasoning item id (rs_
) without the required next item, and it rejects the whole request with a 400.

What to do

  1. Check your OpenClaw version
  • Please paste: openclaw --version
    This was fixed in OpenClaw 2026.1.10 (changelog: “avoid standalone reasoning items that trigger 400 ‘required following item’”).
  1. Immediate workaround (gets you unstuck right now)
  • In the affected Telegram chat, send /new as a standalone message.
    That starts a fresh session id so the “bad” prior transcript won’t be replayed.
  1. If you’re already on a recent version and it still happens
  • Paste the relevant gateway log snippet around the failure (or openclaw status --all output, redacting tokens).
  • Also tell me which model/provider you’re using (just “OpenAI via Responses API” vs something else is enough).

Docs refs:

worn kelp
#

I checked everything but same problem

delicate orioleBOT
#

That 400 is very specific to OpenAI Responses transcript formatting: the Gateway is sending a history that contains a reasoning item id (rs_
) that isn’t followed by the required next item. So if it keeps happening, we need to figure out (a) are you actually on a build that contains the fix, and (b) is /new truly resetting the session you’re talking in.

Please paste these 3 things (they’ll make the cause obvious):

  1. Your version
openclaw --version
  1. Model/API mode being used
  • In the same Telegram chat where it fails, run:
    • /model status
    • (and optionally) /think (just to see what it’s set to)
  1. A log snippet of the failing request
# 
openclaw logs --follow

Then reproduce the error once, and paste the lines around the 400.

Two quick gotchas to double-check (common “I tried /new but nothing changed” cases)

  • Did /new reply with a little “reset/hello” confirmation?
    If it didn’t, then /new likely didn’t apply (commands not being processed/authorized), so the poisoned history remains.

  • Telegram group command routing: in groups, try sending:

    • /new@<your_bot_username>
      (Telegram sometimes routes commands weirdly otherwise, depending on client + group settings.)
#

Temporary workaround (if you need it working immediately)

If /model status shows you’re using openai-responses, switch to a provider/API mode that uses chat/completions (or a non-OpenAI provider) until we get the version + logs. This avoids the Responses “reasoning item ordering” strictness entirely.

Send the 3 outputs above and I’ll tell you the exact next step (it’s usually either “you’re on an older build / gateway didn’t restart” or “your provider is set to openai-responses and the session reset didn’t actually happen”).

polar fox
#

I have the same issue since I've updated to 2026.2.15 . The problem is with "primary": "openai/gpt-5-mini", but not in "openai/gpt-4.1-mini"

delicate orioleBOT
#

Yep — this is a known regression in OpenClaw v2026.2.15 with OpenAI “reasoning” models (GPT‑5*), and it matches that exact error:

400 Item 'rs_...' of type 'reasoning' was provided without its required following item.

Why it only happens with openai/gpt-5-mini: GPT‑5 models emit reasoning items in the Responses API stream, and OpenClaw ends up persisting/replaying a dangling reasoning item in the conversation history after a couple turns. OpenAI rejects that history unless each reasoning item is immediately followed by a message/tool item.
openai/gpt-4.1-mini doesn’t produce those reasoning items the same way → so it “works”.

Refs (same bug being tracked):

Workarounds right now

  • Keep primary on openai/gpt-4.1-mini (what you already found).
  • If you really need GPT‑5: start a fresh session when it breaks (temporary; it’ll recur).
  • If this started exactly with 2026.2.15 for you: downgrade to the last working OpenClaw version until a patch release lands.