fix(subagent): reliable gateway tool-loop for non-Anthropic providers (persist tool-result messages + JSON-normalize outputs)

[rafaelreis-r]

Fixes #2256.

The gateway-native subagent loop (agent.use_gateway_loop=true, mandatory for non-Anthropic providers) couldn't reliably complete a multi-turn tool conversation. Two independent bugs combined to make subagent jobs loop on retries forever. Both surface as AI SDK v6 errors: Invalid prompt: The messages do not match the ModelMessage[] schema and Tool results are missing for tool calls <ids>. The Anthropic-direct path is unaffected.

Commit 1 — persist tool-result messages; self-heal on resume

toolLoop() fed each turn's tool results back as a role:'user' message but never persisted it (void userMessageIdx), and runSubagentViaGateway discarded ToolLoopResult.messages. So subagent_messages had gaps at every even message_idx. On any resume, loadPriorMessages rebuilt a history of adjacent assistant turns with dangling tool-calls, which AI SDK v6 rejects — before the replayState.priorTools reconciler (which only runs inside tool-dispatch, after chat()) ever gets a chance. The job then looped to max_attempts. The legacy Anthropic-direct path always persisted this message.

• toolLoop: new onToolResultMessage callback, fired before the in-memory push (write-before-use).
• runSubagentViaGateway: wires it to persist the tool-result message; and rebuildReplayHistory() self-heals already-corrupted jobs by reconstructing each missing tool-result message from settled subagent_tool_executions rows (matched by provider toolCallId) and re-persisting it. Conservative: a still-pending/absent exec leaves the turn un-healed rather than fabricating data. nextMessageIdx now derives from max(known idx)+1 so a healed gap can't collide under the unique (job_id, message_idx) constraint.

Commit 2 — JSON-normalize tool-result outputs

AI SDK v6 validates each tool-result json output against a strict JSONValue schema. node-postgres returns timestamptz as JS Date, so brain_get_page/brain_list_pages rows carry Date-typed updated_at/created_at; a raw Date (also undefined/bigint) fails the ModelMessage union the turn after such a tool runs. Replay masked it (the jsonb round-trip had already stringified the Date).

• toModelMessages: normalize the json value via toJsonValue() (JSON round-trip: Date→ISO, undefined dropped, non-serializable→string), plus a non-throwing safeStringify() for the error-text branch. Persisted blocks are unaffected (write path already JSON-stringifies, so stored and sent forms now agree).

Tests

• test/e2e/subagent-gateway-resume.test.ts — forward persistence (contiguous message_idx, no gaps) + self-healing resume from a pre-fix corrupted job without re-executing the tool.
• test/gateway-model-messages.test.ts — Date→ISO normalization, undefined drop, and a real generateText validation accepting a Date-bearing tool output (no ModelMessage rejection).
• Existing subagent-gateway-path and cross-provider subagent-crash-replay-multi-provider suites stay green (33 pass on this branch).

Validation

Deployed against a live llama-server:qwen36b dream backlog where 23/23 multi-turn jobs were stuck looping. After the fix, jobs that previously failed every retry run the full multi-turn synthesis to completion with contiguous message_idx (verified one job: 16 messages, idx 0–15, real synthesis output; the rest drained the same way). The only non-completions were pre-existing wall-clock timeout deaths on the heaviest transcripts that had already exhausted most of their retry budget on the old bug — not a regression.