Context window
- Requests asked for the model's full max_output, but providers count
input + output against the window: long chats with large-output models
(DeepSeek V4.1 Flash, 384k) failed with "input too long" before automatic
compaction (85 %) kicked in. max_tokens is now what the window leaves
(estimated input = last usage + messages since), the chat is compacted when
less than 32k would remain for an answer, and a context-overflow error
compacts and retries once.
- ModelConfig.max_input (catalogue limit.input, editable in the provider
settings) lowers the compaction threshold for models with an input limit.
Enforced reviews (modes with refine.rounds > 0, refine.after_todos default on)
- Every todo marked completed is reviewed against the work since the last
review; with findings it is reopened and the feedback goes to the agent
(max. rounds per todo). Passed reviews leave a visible note.
- Ending the turn with open todos sends the agent back to work while it makes
progress (max. 10 reminders in a row without progress); such modes may take
up to 2,000 steps per turn. Reaching the step limit is reported.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Richard ZinkandClaude Opus 5.5
bd896794fd