fix(coding-agent): remove blanket RPC timeouts, reject on real boundaries

[snimu]

What was wrong

Every RPC call had a hidden 30-second deadline (10 minutes for refine), and waiting for the agent to finish had a hidden 60-second one. Long-but-healthy commands — like a Bash call that legitimately takes minutes — got cancelled on the client while they kept running on the server. And because several timers overlapped, the same operation could be bounded three different ways.

The fix

• Removed all the default deadlines. Callers can still pass a timeout explicitly if they want one.
• Instead of timers guessing, failures now come from the places that actually know: the child process dying, its output stream closing, or an explicit stop(). Anything still waiting gets a clear error at that moment.
• Replaced the test that asserted the old 10-minute refine deadline with three tests for what actually matters: a long command finishes, a long turn finishes, and everything pending fails cleanly when the process dies.

How it's verified

Type checks, lint, focused suites, and the repo-wide check pass. The process-death tests were initially flaky (two shutdown events race each other); the assertion was fixed and then run 11 times in a row without failure. Written and independently reviewed by two different models before opening this PR.

Note: intentionally no Linear ticket for this cleanup stack, so that check stays red.

---

Note

Medium Risk
Behavior change: callers that relied on implicit RPC/idle timeouts now wait indefinitely until transport failure unless they pass timeout; lifecycle fixes are complex but heavily tested.

Overview
Removes default client-side deadlines on RpcClient RPC calls and agent-turn waits so long-running bash, refine, and multi-minute turns are not cancelled while the server is still healthy. send() no longer applies a 30s (or 10-minute refine) timer; waitForIdle, collectEvents, and promptAndWait only time out when callers pass an explicit timeout.

Pending work now fails on real transport boundaries instead of overlapping timers. A latched transportError plus failPendingOperations() rejects in-flight commands and event waiters when the child errors, stdout closes, the process exits (with a short stdout drain so last responses can still resolve), stop(), or start() replaces a generation. Handlers are scoped per child so restarts and grandchild-held stdio pipes cannot let stale output resolve new-session waiters.

prompt() awaits the RPC response and surfaces command failures via getData. promptAndWait uses cancellable startEventCollection() so prompt rejection does not leave idle waiters hanging.

Tests replace the refine-timeout assertion with a hanging fixture and suites covering no default timeout, process death, restart isolation, and ghost events.

^{Reviewed by Cursor Bugbot for commit 6bf726d. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Remove blanket RPC timeouts in RpcClient, reject on transport failure boundaries

• Removes the per-request 30s timeout and the 10-minute REFINE_REQUEST_TIMEOUT_MS constant; send() and refine() now wait indefinitely until a response arrives or the transport fails.
• Adds failPendingOperations which latches the first transport error and rejects all pending requests and event waiters when the child process exits, stdout closes, or an error occurs.
• prompt() now awaits the command response and throws on rejection instead of being fire-and-forget.
• start() cancels prior pending operations before spawning a new child; stop() always resolves within ~1s via SIGKILL fallback even when a grandchild holds stdio pipes open.
• Behavioral Change: waitForIdle(), collectEvents(), and promptAndWait() no longer time out by default and reject on transport failure; callers relying on the old 30s/10min timeouts must pass an explicit timeout.

^{Macroscope summarized 6bf726d.}

Linear: ENG-5647

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit f1a560e. Configure here.}

[snimu]

Folded into #1854: that branch already contains this PR's commits (the two changes share code and were sequenced as a merge-after chain). Merging #1854 lands both; per snimu's preference the chain is collapsed into one merge unit. The review history here remains valid for the contained commits.