[Model behavior] Rushes tasks to a finish and ships multiple bugs instead of completing them correctly

In a long working session the model repeatedly pushed tasks to a "done" state quickly and left defects the user then had to find.

Examples from one session:

• A size-reduction change to a container image removed a standard-library component that downstream users needed; it shipped in a release and was reported as a breaking change.
• A newly added emulated service blocked the API call that creates an instance until a container was running; the first call exceeded the SDK timeout, the SDK retried, and the retry failed with an "already exists" error.
• A follow-up fix for that call was delivered while an identical blocking pattern remained in a sibling operation; the user had to ask for it.
• A release tag was prepared without the model checking that the branch it described had been pulled.

Repro steps

1. Give the model a multi-part task in a large repo (implement a fix, prepare a release, review PRs) with instructions requiring verification before claiming done.
2. Let it work through several such tasks in one session.
3. Observe "done" reports followed by defects in the same surface that a full check would have caught.

Expected

Each task finished completely, the surrounding code checked for the same defect class, and verification run before reporting done.

Actual

Fast "done" reports with defects left for the user to discover.

Environment

Claude Code CLI, macOS, Opus model, long sessions with context compaction.

Related

#101574, #101576, #101577