01: Stream Once

One provider generation pass, wrapped with a rolling AbortController SDK timeout.

File: src/utils/chat/toolLoop.ts:142-195

Mission

Call provider.streamToDiscord(...) with the current accumulated context and tool history, and race the result against a configurable SDK timeout. The timeout is rolling: it resets on every onStreamProgress heartbeat, so a long but active stream is not killed — only a truly stalled one is. Returns a StreamResult describing how the generation ended.

Input

• params: ToolLoopParams — full loop context (provider, config, ChatTurnContext).
• accumulatedModelParts: Array<Record<string, unknown>> — provider-native model turn parts accumulated across prior iterations of the tool loop. Empty on the first iteration; grows as each tool call appends its model response. Passed by reference and read (not written) inside streamOnce.
• functionHistory: ToolHistoryEntry[] — paired call/response records from prior tool dispatches. Passed to the provider so it can continue the multi-turn tool conversation. Empty on the first call.

Output

Promise<StreamResult> — defined in src/types/provider/interfaces.ts. The status field drives the outer loop’s switch:

status	Meaning
"completed"	Provider finished; accumulatedText carries the final response
"error"	Provider threw a non-timeout error
"timeout"	SDK call exceeded STREAM_SDK_CALL_TIMEOUT_MS with no heartbeat
"empty_response"	Provider returned with no text and no tool call
"stopped_by_user"	User triggered /stop while streaming
"follow_up_interrupt"	A follow-up message arrived; caller should yield
"function_call"	Provider requested a tool call; data carries the call payload

Side effects

• Sets params.context.streamingContext.abortSignal to a fresh AbortController.signal before each call. Provider adapters consume this signal to abort in-flight HTTP requests when the timeout fires.
• Sets params.context.streamingContext.onStreamProgress to the refreshTimeout callback before the call and resets it to undefined in the finally block. Provider adapters call this on each token delivery to prevent the timeout from firing on active streams.
• Registers killStream on the channel lock entry via setChannelStreamKill(channelId, killStream). killStream is a unified callback that both calls abortController.abort() and rejects the Promise.race — ensuring the HTTP request is cancelled and the race unblocks simultaneously. This is what /bot kill triggers via forceKillChannelStream.
• Clears the timeout and the kill registration (clearTimeout, setChannelStreamKill(channelId, null)) in the finally block regardless of success or error.
• Derives the replyToMessage argument passed to the provider. Queued turns (isFromQueue) normally reply to their trigger message so the response renders as a Discord reply. Scene turns are the exception: every queued scene persona job shares the same trigger message, so replying would make all of them render “replying to” one message. When incoming.sceneTurn is set, replyToMessage is forced to undefined so the generated scene reads as a free-standing dialogue.

Invariants

After this stage runs:

• params.context.streamingContext.onStreamProgress is undefined — the heartbeat reference is always cleaned up.
• The channel lock’s activeStreamKill is null — the kill callback is always deregistered in finally.
• If the result status is "timeout", it originated from the SDK-call timeout race (error message prefix "SDK_CALL_TIMEOUT:"), not from a provider-specific timeout mechanism.
• Errors that are not SDK timeouts are re-thrown to the caller (runToolLoop), which does not catch them — they propagate to runGenerationTurn’s outer catch.

Extension points

Surface	Plugin-relevance
provider.streamToDiscord(...) call	The provider contract is the seam — see provider pipeline
STREAM_SDK_CALL_TIMEOUT_MS / rolling onStreamProgress	Internal — timeout behavior is an operational concern, not plugin-relevant
params.context.streamingContext.abortSignal	Internal — consumed by provider adapters only

Configuration

Env var	Default	Minimum	Purpose
STREAM_SDK_CALL_TIMEOUT_MS	120000 (2 min)	10000 (10 s)	Idle timeout for one provider call; resets on each heartbeat

Related docs

• Provider streaming contract: → provider pipeline
• Tool-loop coordinator: → README.md