Error with "redacted thinking" after a while causing the need for a new session. #173
Description
⚙️ Status
Web: not linked — run clawdis login to scan the QR.
Agent: embedded pi • anthropic/claude-opus-4-5
Workspace: ~/clawd
Context: 71k/200k (36%)
Session: main • scope per-sender • updated just now • store ~/.clawdis/sessions/sessions.json
Options: thinking=low | verbose=off (set with /think <level>, /verbose on|off, /model <id>)
Shortcuts: /new reset | /restart relink
After the agent runs for a while (and I suppose some heartbeats happen on haiku 4.5 for a few days) I get an error about thinking
LLM request rejected: messages.147.content.1: thinking or redacted_thinking blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response.
The only workaround here for me is to reset the session. I believe this is related to the extended thinking feature in the anthropic API.
Rata (my clawdis's chosen persona) said that (fwiw, take with a grain of salt)
The real fix is in the conversation history handling - those thinking blocks shouldn't be getting modified between requests in the first place. Something's mangling them during storage/replay.