Problem

Every subagent routed through the gateway tool-loop (agent.use_gateway_loop, i.e. any non-Anthropic provider) that emits parallel tool calls permanently fails on retry with:

[chat(<provider>)] Tool results are missing for tool calls call_…, call_…, call_…

i.e. AI SDK v6's AI_MissingToolResultsError. In production on GLM (glm-coding) this killed cron/agent jobs the day after the gateway loop took over — company-pulse and every dream-synthesize subagent that fanned out parallel tool calls died, and the retries compounded the failure.

Root cause

The gateway loop persists tool executions to subagent_tool_executions, not as subagent_messages rows. On replay, runSubagentViaGateway rebuilt the prior conversation with a naive 1:1 map:

const priorChatMessages = priorMessages.map(m => ({ role: m.role, content: adaptContentBlocksToChatBlocks(m.content_blocks) }));

That yields assistant tool-call turns with no following tool-result message. The loop seeds messages from replayState.priorMessages, so the next generateText() sees unanswered tool-calls and v6 rejects the prompt. A single crashed turn with N parallel calls drops all N tool-results at once.

Fix

• Extract a pure, unit-testable reconstructReplayMessages(priorMessages, priorToolExecs) that re-inserts a tool-result message after each prior assistant tool-call turn, pairing every call with its persisted outcome (keyed by message_idx + provider tool_use_id). A call with no completed execution (crash before the result was persisted) gets a synthesized error result so the turn still validates and the model recovers next turn.
• Derive nextMessageIdx, nextTurnIdx, and the initialMessages seed-condition from the persisted-message count (priorMessages), not priorChatMessages.length — the latter now also counts the re-inserted tool-result messages (which are not their own subagent_messages rows), so the loop's messageIdx counter would otherwise skip.

Provider-neutral; only reachable on the gateway path.

Why the existing tests didn't catch it

test/e2e/subagent-crash-replay-multi-provider.test.ts stubs the chat transport (__setChatTransportForTests), so it short-circuits before generateText runs and never exercised the ModelMessage conversion — the same blind spot called out in test/ai/gateway-tools-schema.test.ts.

Testing

New test/subagent-replay-toolresult-repair.test.ts (4 tests, no network):

• re-pairs every parallel tool-call with its persisted tool-result;
• no spurious tool-result message for non-tool turns;
• synthesizes error results for failed / missing executions;
• drives the real AI SDK v6 generateText (via MockLanguageModelV3) and asserts the re-paired history validates while the naive 1:1 map throws.

Regression: subagent-crash-replay-multi-provider (13), gateway-tools-schema (3), gateway-tool-loop (7) all still pass.

Context: this is the upstream reconciliation of a long-running local fork patch (the f9c22f7 "C-3" the fork notes said should land upstream alongside #1708/#868).

🤖 Generated with Claude Code