Affected version: [email protected] (also present in 2026.5.27)

Summary

When an OpenAI-compatible provider streams a structured tool call but reports finish_reason: stop (instead of tool_calls), OpenClaw maps the turns stopReason to stop, never executes the pending tool call, and the run terminates with terminalError: non_deliverable_terminal_turn. The result is silently dropped: no assistant reply and no error message are delivered to the user.

The streaming response assembler already guards the opposite inconsistency (stopReason === toolUse but no tool-call blocks -> downgrade to stop), but the symmetric case (stopReason === stop but tool-call blocks are present) is not handled.

Impact

• User-visible silence: the agent appears to not answer. The turn ends in status: error but, because it is classified non-deliverable, nothing reaches the channel.
• replaySafe: no for this class, so there is no automatic retry.
• Observed intermittently in production (~12 incomplete-turn terminations over 3 days) against a self-hosted vLLM (Qwen/Qwen3.6-27B-FP8, openai-completions API).

Environment

Root cause

1. finish_reason is mapped purely by string in openai-transport-stream.ts -> mapStopReason(). finish_reason: stop always maps to stopReason: stop, regardless of whether the assembled message contains tool-call blocks.

2. The stream finalizer guards only one direction:

const hasToolCalls = output.content.some((block) => block.type === "toolCall");
if (output.stopReason === "toolUse" && !hasToolCalls) output.stopReason = "stop";
// <- missing: the symmetric stop but DOES have tool calls case

3. Downstream classification in attempt-trajectory-status.ts -> resolveAttemptTrajectoryTerminal(): With stopReason === stop, no visible assistant text and no delivery evidence, the turn falls through to non_deliverable_terminal_turn.

Proposed fix

Add the symmetric guard in the streaming finalizer of openai-transport-stream.ts:

const hasToolCalls = output.content.some((block) => block.type === "toolCall");
if (output.stopReason === "toolUse" && !hasToolCalls) output.stopReason = "stop";
if (output.stopReason === "stop" && hasToolCalls) output.stopReason = "toolUse"; // NEW

Why this is correct:

• Mirrors the existing guard at the same point after finishAllToolCallBlocks()
• If a provider emits structured tool calls but reports finish_reason: stop, the turn is treated as toolUse and the runtime executes the tool and continues
• Only promotes stop -> toolUse when tool-call blocks are actually present; normal text completions unaffected; does not touch length (truncated args must remain truncation)

Optional hardening

If a non-streaming assembly path exists, it would benefit from the same hasToolCalls && stopReason === stop check.