[Bug]: LLM consistently doesn't see its tool call outputs in next iterations

[alexhooketh]

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• I am using the latest version of Moltis
• If this happened during a chat session, I included as much full session context as possible and redacted secrets

What happened?

noticing this behavior in pretty much every chat — the LLM seems to not be able to see its tool call outputs when i send a new message to the chat. this makes it confused and think it has hallucinated everything above and i have to verbally calm it down. not just that, when it becomes convinced that it did produce valid data with (as it sees) no tool calls earlier, it starts outright hallucinating the info confidently, rendering chats unusable after just a few messages
i'm using telegram

Expected behavior

pretty sure that openclaw does include tool calls in the chat history, because i've never encountered this issue (or even the agent saying that it has hallucinated anything, really) on openclaw. i'd expect this behavior to be the same. i believe that it's either a bug or an attempt to save context usage — if the latter, it's certainly an unsuccessful one

Steps to reproduce

run a moltis agent on telegram. i use gpt-5.2 via codex oauth
ask it to do anything tool-heavy
signs of it not seeing its own tool call outputs are obvious after just a few messages

Did this happen during a chat session?

None

Chat session context (if applicable)

Error messages / logs

i think that it's rather undesired behavior than a bug

Is this a regression?

No, this never worked

Last working version

No response

Moltis version

0.10.11

Component

Agent / LLM providers

Install method

Pre-built binary

Operating system

Ubuntu / Debian

Additional context

No response

[lostmsu]

Yes, this is very broken behavior, which has terrible consequences:

1. agents are extremely dumb
2. this burns through tokens because resubmitted conversation does not match the original one, so prefix caching does not work

[penso]

Looking at this