Ctrl-C during streaming ended the run without showing the interrupt menu
A single ^C during an Opus reasoning stream ended the turn immediately. No interrupt menu appeared, no message was printed, and the process exited 0. Not reproduced since.
What the trace log shows
Thinking deltas were still arriving 1.6s before the signal, so the stream was live:
{
"timestamp": "2026-08-11T10:23:49.150307Z",
"level": "TRACE",
"fields": {
"message": "Received event from Anthropic API.",
"event": "...thinking_delta..."
}
}
{
"timestamp": "2026-08-11T10:23:50.788483Z",
"level": "INFO",
"fields": {
"message": "Signal received.",
"signal": "SIGINT"
}
}
{
"timestamp": "2026-08-11T10:23:50.788522Z",
"level": "DEBUG",
"fields": {
"message": "Routed OS signal.",
"signal": "Interrupt",
"routed": "Handler"
}
}
{
"timestamp": "2026-08-11T10:23:50.788671Z",
"level": "INFO",
"fields": {
"message": "Interrupt received during streaming."
}
}
{
"timestamp": "2026-08-11T10:23:50.805940Z",
"level": "INFO",
"fields": {
"message": "Flushed conversation to disk.",
"id": "jp-c17860012287"
}
}
{
"timestamp": "2026-08-11T10:23:50.849953Z",
"level": "DEBUG",
"fields": {
"message": "RunningService dropped..."
}
}The router picked the right handler and handle_streaming_interrupt (crates/jp_cli/src/cmd/query/interrupt/signals.rs:70) ran. 17ms later the conversation was flushed and teardown began, so InterruptHandler::handle_streaming_interrupt returned a decision without ever blocking for input.
On the terminal, the buffered reasoning appeared right after the echoed ^C (consistent with the flush_renderer() + flush_instant() at the top of that function) and nothing else. No menu, no Interrupted.
The evidence does not add up
Two ways out of that function in 17ms, and neither fits cleanly.
config.action != Prompt (handler.rs:192) skips inline_select entirely and hardcodes the choice. stop and abort both end the turn with Ok(()), which matches the exit status. But the action was not set anywhere: not in ~/Library/Application Support/jp/config.toml, not in the user-global config/ directory, not in a user-workspace config (no such file for workspace otvo8), not in JP_CFG_INTERRUPT_*, and not as a config_delta on the conversation. So it should have resolved to the Prompt default.
inline_select returning Err maps to InterruptAction::Escalate (handler.rs:215), which cancels the shutdown token and returns cmd::Error::interrupted(). That propagates unchanged through query.rs:942 and run_inner's select, so it should have printed Interrupted on stderr and exited 130. It did neither.
One of those two readings has to give. The likeliest weak link is the exit status, which was recalled from a shell prompt rather than read from $?. If it was really 130, this is inline_select failing instantly and the question becomes why.
Why it can't be timing
The menu decision does not consult stream state. stream_alive is only used inside the 'c' branch to pick Resume over Continue; it never suppresses the menu. A ^C landing between the thinking block and the first text delta, or after the stream died, still prompts.
Next time
Err(_) at handler.rs:215 and handler.rs:296 discards the InquireError, which is exactly the value that would have settled this. Same for the Ok(ReplyOutcome::Cancelled) | Err(_) arm in collect_reply_inline. Nothing logs the resolved interrupt.streaming.action either, so ruling the config in or out took five greps across four layers instead of one grep in the trace log.
Adding that logging is the immediate work; the bug itself stays open until it reproduces with the extra data.
Related
RFD 060 (Config Explain) would have answered the config half directly.
Comments
- From: jp
- Date: 2026-08-11T11:15:05Z
Logging is in place.
handler.rs now logs the resolved action (Handling streaming interrupt. with action and menu), and log_prompt_failure records the InquireError that the menu call sites previously threw away as Err(_).
signals.rs logs the chosen InterruptAction and the returned result.
Still open until it reproduces.
- From: jp
- Date: 2026-08-17T16:35:48Z
Fixed one proven cause of this symptom, though not the evidence recorded above.
The escalation ladder was counting presses it had already answered. EscalationState::bump only reset on elapsed time, so answering a menu and interrupting again inside the 2s cooldown counted as press two and routed straight to Shutdown, bypassing the handler stack — no menu, no message. A delivered press is now an InterruptNotice the handler resolves: handled() clears the count, decline() and dropping leave it intact. Only unanswered presses escalate.
Also split a user-cancelled menu from one that could not run. The latter returns PromptFailed and leaves the press on the ladder rather than escalating on a decision nobody made.
Confirmed live: eight presses, five inside the cooldown of an answered one, all reached the handler.
Staying open because the traces in the description showed routed=Handler with a ~20ms return — a different signature from this bug — and those files are gone.