feat(agent): turn-level retry with idle timeout, plus withToolRetry helper

[robert-j-y]

Summary

Fusion benchmark failures traced to mid-loop stream deaths: a turn's SSE stream opens fine, emits ~0–5 tokens, then either ends with no terminal event (Follow-up stream ended without a completed response) or hangs silently forever. Today the whole callModel promise rejects and callers can only restart from turn 0 — re-running every tool call and losing all gathered context.

• retryTurn option on callModel — { limit, isRetryable, backoffMs, idleTimeoutMs }. Wraps each turn's send+consume (initial, follow-up, and final-response turns) and re-sends just the failed turn's request with the accumulated conversation state intact. Typed errors (TurnStreamEndedError, TurnResponseFailedError, TurnIdleTimeoutError) drive the default retryable classification (5xx/transport/stream-death yes; 4xx account-state no).
• idleTimeoutMs — converts silent stream hangs (stream never closes, nothing throws) into retryable turn failures instead of relying on an external inactivity guard killing the request.
• turn.retry event — emitted on the response stream before each retry attempt so consumers can observe retries.
• withToolRetry helper — separately wraps a tool's execute with bounded retries so transient tool throws are absorbed instead of being serialized as {"error": ...} results that burn a model turn (per Matt's suggestion).

Includes changeset for a minor release (0.8.0). Note: branch builds on the two unpushed atomicity commits (ce00bf9, 7b6704a).

Test plan

• New unit tests: tests/unit/turn-retry.test.ts (follow-up/initial turn retries, idle timeout, retry limits, custom isRetryable, backoff, turn.retry event semantics)
• New unit tests: tests/unit/tool-retry.test.ts (regular + generator tools, limit exhaustion, isRetryable, onRetry, type preservation)
• Full unit suite, typecheck, lint green
• Wire into fusion inner calls in openrouter-web after 0.8.0 publish (prepared behind a TODO at the callModel site)

Made with Cursor

[robert-j-y]

Parking this for now per team discussion — we'll handle recovery at the application layer (fusion/server-tools in openrouter-web) until the comments-based stream-liveness approach is designed for the SDK. Branch stays up for reference.