[loosheng]

Summary

Cursor can leave a Run's stream open after some thinking with nothing more to come: no text, no call, no count, no TurnEnded. Neither path ends such a Run:

• stateless runOnce / decode: ends only on an end frame, TurnEnded, a complete count or EOF; the plugin's 5 s heartbeats keep the stream alive. No timer at all.
• retained NativeSession.pump: batchWait starts only at a count or a call, so before either only the 15 min lifetime is left.

In yetone/magpie#2672 two composer-2.5 turns waited 305 s until omp gave up (magpie: 499, usage: null). They were on the stateless path: their call ids were Cursor's own tool_…, not call_cursor_….

every frame of a Run, both paths
  → Quiet: time spent waiting on Cursor's next frame
      no frame at all for idle (60 s)                 → 502 "Cursor went quiet: …"
      only InteractionUpdate.heartbeat for quiet (240 s) → 502 "Cursor went quiet: …"
      retained Run in awaiting-results                → paused (resultWait bounds it, as before)

• Each frame resets idle, so a step whose calls keep coming still seals however long it takes in all (cursor: "Cursor did not complete its declared native tool batch" (502) when a large parallel tool step stalls under batchWait=30s #155); batchWait is unchanged.
• quiet is under magpie's 5 min keepalive, so a Run that gets only Cursor's heartbeats still ends before the agent's own stall timer.
• The error names how long, and the last kind of message Cursor sent (thinking, tool call, …): the next report says what Cursor did.
• Stateless: a quiet Run is an error even after some calls, not the EOF-with-calls success.

@cursor/sdk 1.0.35 ends a Run the same way (stall detector, 30 s without a message, then retries; dist/cjs/689.js). Its server-heartbeat path resets its meaningful-activity clock too (reset("inbound_message","heartbeat"), third arg defaults true), so Cursor's heartbeats would hold such a Run open there; here they are counted apart.

Evidence

• Before: the 8 new cases against unchanged index.mjs / native-session.mjs: 6 fail, 2 pass. Failing: thinking then silence (native, stateless); Cursor heartbeats only (native, stateless); a call with no count then silence (stateless); a call then silence before batchWait (native). Passing either way: a step of calls each within idle; a retained Run parked past idle continues same-run.
  After: all 8 pass; HOME=$(mktemp -d) bun test: 2452 pass, 0 fail (213 files); bun scripts/check.mjs ok.
• Live, a Cursor Pro account via cursor-agent's token, loader.fetch with one read_file tool: auto native and stateless, a call then "Pelican", 200 each; composer-2.5 native and stateless, 200; auto native with idle at 3 s and the continuation sent after 8 s parked: same-run, 200.
• Not reproduced: Cursor going quiet for real. Whether it then sends nothing or its own heartbeats is unknown; both are covered by tests.

Merge Danger

Door: two-way

Revert restores the old behavior; no schema, config or stored state.

Blast Radius: Cursor

A Run where Cursor is legitimately silent for over 60 s (no message, not even a heartbeat) between frames, or over 4 min with heartbeats only, now fails with 502 instead of waiting. Time parked for the caller's tool results never counts.

[yetone]

Reviewed at cbd5d3b (on main df55294). Context: yetone/magpie#2672.

The design is right and the tests are good:

• idle is 60s with no frame and quiet is 240s with heartbeats only. Both are under magpie's 5 min keepalive ceiling (keepaliveLongest, internal/gateway/gateway.go).
• The timers pause while a Run is awaiting results and resume after the submit.
• Both the stateless path and the retained path are wrapped.
• The errors are clear, and 0.2.21 is the right bump (npm has 0.2.20).

I broke each part by hand and the tests went red each time:

Change	Tests failed
stateless path unwrapped	3
no pause at seal	2
heartbeat counted as a frame	2
idle not reset per frame	3, including "a step whose calls each come within idle outlasts idle in all", which covers #155

bun test in packages/cursor passes (Bun 1.4.2).

Blocking: Quiet.frames keeps every frame alive for the life of the stream

const gone = new Promise((r) => { this.gone = r })
...
r = await Promise.race([next, gone])

gone is created once and stays pending until the stream ends. Each Promise.race adds a reaction to gone, and that reaction holds the race's result promise, which holds the frame it resolved with. None of it can be collected until gone settles. A long Cursor run streams many MB of frames, so memory grows with the whole turn.

Repro: run a script that sends 20,000 frames of 50KB through new Quiet(NATIVE_LIMITS, () => {}).frames(src) and prints process.memoryUsage().rss at the end.

• At this head: 1376 MB RSS.
• With the fix below: 63 MB.

Fix (verified; the suite still passes, and the heartbeat test passed 6 of 6 repeats):

r = await Promise.race([next, new Promise((w) => { this.gone = w })])

This makes a fresh wake promise each iteration and drops the long-lived gone. fire and pause still call this.gone, which is now the current iteration's resolver.

Minor

• The heartbeat e2e test has a 350ms lower bound. Under a load average around 400 it took 282ms once. Give it more room on that side, or wait for the state instead of measuring the time.

Not tried: a real Cursor account, Windows.

[loosheng]

Thanks. Both fixed in 6a17d61.

Blocking: Quiet.frames held every frame

Reproduced at cbd5d3b: 20,000 frames of 50 KB through Quiet.frames → 1458 MB RSS. With your fix (a fresh wake per read; fire/pause wake the current one, cleared after each read) → 108 MB.

Added a regression test that doesn't depend on RSS: 200 frames of 64 KB through Quiet.frames with the stream kept open, then Bun.gc(true) and count the frames still reachable via WeakRef. At cbd5d3b it fails (200 held); now it passes (at most 2, the generator's own last value).

Minor: heartbeat test lower bound

Dropped both timing bounds. The test now checks that the Run ended on the heartbeats-only branch (N heartbeats and nothing else, N > 0), which only quiet reaches; idle alone would say no message for. Ran it 8 times with 2× cores of busy loops: 8 of 8 pass.

Run at this head

• HOME=$(mktemp -d) bun test: 2453 pass, 0 fail (213 files).
• bun scripts/check.mjs: every package ok.

[yetone]

Merged at the reviewed head 6a17d61 as cf1873e. Thanks for the quick turnaround.

Re-reviewed every line at 6a17d61. The memory fix holds: 20,000 frames of 50 KB through Quiet.frames, stream left open, end at 95 MB RSS (1376 MB at cbd5d3b). packages/cursor bun test: 288 pass, 0 fail; bun scripts/check.mjs ok (Bun 1.4.2, sandbox HOME).

Broken by hand: the long-lived gone put back, heartbeats counted as frames, and the stateless path unwrapped each turned tests red. One break turned nothing red: if resume() left the Run paused, a retained Run that went quiet on its second model call would wait for lifetime again. I added a test for that in c9a0215 (a continued Run that goes quiet after the caller's results). It is red with that break and green at your head.

Not tried on my side: a real Cursor account, Windows. 0.2.21 is publishing now. magpie's Cursor mover min will be raised so moved users get it.