fix(agent): retry empty final turns instead of silently abandoning the task

[snimu]

Providers occasionally end a turn with a normal stop reason and zero usable content: no text, no tool calls, sometimes a thinking-only message (observed and raw-SSE-confirmed for moonshotai/kimi-k3 via Prime Inference: single-chunk stream, no content, usage.output=0; ~1/10 raw probes). The agent loop treated any no-tool-call turn as completion, so the agent silently abandoned its task: print mode printed nothing and exited 0; interactive mode thought for minutes and then did nothing.

The fix

• streamAssistantResponse now retries an empty final turn silently, up to 3 attempts. The retry happens before message_end is emitted, which is the durability edge: discarded attempts are popped from the request context and never reach session persistence, usage accounting, or transcript consumers.
• Empty means: no tool calls and no non-whitespace text (thinking does not count). error, aborted, and length stops are excluded - they are signals of their own, and an identical resend cannot help.
• Context-overflow shapes pass through untouched (!isContextOverflow(message, contextWindow) guards the retry, reusing the existing detector): a silent-overflow turn is empty by definition, and retrying it would burn full-context requests and then mask the stop reason that auto-compaction recovery keys on.
• After 3 consecutive empty turns, the final message becomes a standard error turn (Model returned an empty response ... 3 times in a row) and flows the existing error path: print mode exits non-zero with the message on stderr; session-level retry/backoff applies as for any other turn error.
• RLM children: a child whose turn ends in an error now delivers RLM child <name> failed: <error> to its parent instead of the misleading "completed without sending a reply" notice. (Found while wiring exhaustion propagation: graceful error turns resolve promptAndWait, so the parent never saw the error text.)

Tests

Eight new tests in packages/agent (silent retry with clean transcript and exactly one message_end; whitespace-only retry; exhaustion after 3 attempts; no-retry guards for real content, tool calls, aborts, empty length stops, and silent-overflow turns - the last two pin the compaction-recovery contract) plus one in agent-session-recursion (parent receives the child failure message, no terminal notice). All retry-path tests were proven red against the pre-fix loop.

Linear: ENG-5795

---

Note

Medium Risk
Changes core assistant-turn completion and message_end timing for every agent run; behavior is guarded and tested but affects persistence, events, and RLM parent notices.

Overview
Fixes a case where providers return a normal stop with no usable output (thinking-only or whitespace text, no tool calls) and the agent loop treated that as a finished turn, so work stopped with no error.

Agent loop (streamAssistantResponse): detects empty assistant turns (no tool calls and no non-empty text; error / aborted / length stops are not retried). It silently re-requests the model up to 3 times. Failed attempts are popped from context and never get message_end, so they do not hit the transcript or retry context. After three empties in a row, the turn is finalized with stopReason: error and a clear errorMessage. Turns that match isContextOverflow are left alone so compaction recovery still sees them.

RLM: when a child finishes without a parent reply, if the last assistant message has stopReason === 'error' (e.g. empty-response exhaustion), the parent gets rlm_child_failure instead of a completed without reply notice.

Tests cover retry hygiene, exhaustion, no-retry guards, overflow passthrough, and parent failure propagation.

^{Reviewed by Cursor Bugbot for commit 537e79e. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Retry empty final turns up to 3 times in streamAssistantResponse instead of abandoning

• Adds MAX_EMPTY_TURN_ATTEMPTS=3 and isEmptyAssistantTurn helper; empty final turns (no tool calls, no non-whitespace text, stopReason not error/aborted/length) are silently dropped from context and retried up to 3 times, after which the message is finalized with stopReason='error' and an explanatory errorMessage.
• Context-overflow turns are excluded from retry via isContextOverflow.
• Refactors single-attempt streaming into streamAssistantResponseAttempt, which no longer emits message_end; the caller emits it exactly once after retry handling.
• Updates AgentSession.runRlmChild so that when a child run ends with no parent reply but the last assistant message has stopReason='error', the parent receives rlm_child_failure with the child's error message instead of the previous completed_without_reply terminal notice.
• Risk: callers relying on message_end being emitted by the inner attempt function must now depend on the outer streamAssistantResponse; in agent-loop.ts the only caller is updated.

^{Macroscope summarized 537e79e.}