A caller that goes away mid-batch (cancel, display error) left the results of
tools that had already run out of history, so the retry ran them again. Append
the collected results first, then emit the events.
A tool batch counted as committed only once every call had its result, so a
batch cut short was dropped whole and the retry ran tools that had already
had effects. Keep the calls that did report a result: their assistant row
stays as the pairing carrier (a request must answer every tool_calls id),
without the text of the unfinished turn, and the calls without a result are
absent so the retry re-issues them instead of assuming they ran. The kept
rows go to the store right away, and a failed rollback write is not allowed
to mask the failure being rolled back.
Add prek.toml so git commit runs ruff-pre-commit (fix then format) and
pdm-based basedpyright. Document install/install-hook; clear type warnings
that made basedpyright exit non-zero under recommended mode.
After each completed assistant+tool batch, persist and keep that prefix on
failure. Rollback only unfinished assistant/stream suffix. pending_retry
and /retry also apply when history ends with tool results (continue model
without re-executing tools).
Decode and stream reasoning_content as ReasoningDeltaEvent; keep it on
assistant messages. CLI shows dim reasoning, toggles stream/verbose, and
prints full tool results when verbose.
Keep incomplete user messages in memory/DB; roll back only
assistant/tool on failure. /retry and resume of orphan-user sessions
re-run without duplicating the user message.
Tool-loop rounds re-send history, so summed prompt tokens are billing
not context size. Track last_request_usage and label turn/session as
billed; clarify /status and end-of-turn lines.
Max-rounds and destructive-tool prompts use asyncio.to_thread with
cancel paused on the main thread so multi-tool turns no longer look
like user cancel. Hooks may be sync or async.
NCBM