dcdpr / jp / 25887922763

65%

push

github

fix(cli, llm): Prevent duplicate tool-call dispatch on spurious flush (#639)

Two related bugs combined to cause tool calls to be dispatched twice.

On the provider side, both the `cerebras` and `llamacpp` SSE parsers
emitted `Event::flush(1)` twice per stream: once from the
`finish_reason` chunk and again from the `[DONE]` sentinel. A
`message_flushed` boolean gate now ensures the message/structured index
is flushed at most once. A drain of `tool_call_indices` on the `[DONE]`
sentinel is also added as a safety net for streams where the
`finish_reason` chunk never arrives.

On the shell side, tool-call dispatch was inferred by inspecting the
conversation stream's tail after every `Flush` event. Because the stream
tail continued to point at the prior `ToolCallRequest`, any duplicate
flush would re-trigger the preparation pipeline. The fix drives dispatch
off `EventBuilder::handle_flush` directly:
`TurnCoordinator::handle_event` now returns a `HandleEventOutcome`
carrying both the state-machine `action` and a `CommittedEvent`.
`CommittedEvent::ToolCallRequest` is populated only when `handle_flush`
actually commits a new tool call; a no-op flush returns
`CommittedEvent::None`, making the dispatch unambiguous regardless of
provider behaviour.

`handle_llm_event` in `signals.rs` is updated to surface the
`CommittedEvent` alongside the `LoopAction`, and the turn loop switches
from the stream-tail inspection to a direct pattern match on the
committed event. Raw provider events are now captured in test snapshots
so that provider-level misbehaviour such as duplicate flushes remains
visible even when `EventBuilder` filters them out.

---------

Signed-off-by: Jean Mertz <git@jeanmertz.com>

76 of 88 new or added lines in 6 files covered. (86.36%)

26179 of 39235 relevant lines covered (66.72%)

184.0 hits per line