fix(handoff): preserve the current instruction

[sentry-junior]

A handoff summary can incorrectly mark a new request complete. The next model then has only that false summary and can end without doing the requested work.

Keep the current user instruction as a bounded, deterministic part of the handoff checkpoint. The model summary remains separate continuation context.

Requested by David Cramer.

--

View Junior Session [Sentry]

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
junior-docs	Ready	Preview	Aug 26, 2026 8:52pm

[dcramer]

are you confident this is the solution?

[sentry-junior]

mostly yes for this failure mode — medium-high confidence.

what we proved from the failing turn (slack:C0B595QDZLL:1787771533.485339, trace 1be0ec99…):

1. handoff did continue (tool ok → haiku summary → grok resumed)
2. the generated checkpoint said Outstanding Asks: None / “PR complete”
3. that wiped the live instruction (handoff and do a cleanup…)
4. the new model then correctly followed a false “done” summary and emitted [[NO_REPLY]]

so the bug wasn’t “handoff didn’t resume.” it was instruction loss across the replacement boundary.

this fix makes the live user instruction a deterministic, bounded part of the handoff checkpoint, separate from the model summary. that closes the exact hole that caused the silent stop.

not claiming:

• summarizer quality is fixed (it can still be wrong about completion)
• every future “model goes quiet after handoff” is this same root cause
• we re-ran the original production turn end-to-end after the patch

if the summarizer keeps saying “none,” the preserved instruction still forces the next model to see the outstanding ask. that’s why i’m confident this is the right shape for this incident.