/continue previously pushed plan_scan_baseline past every restored
message, so plan_state.is_active() saw an empty slice and hid the card —
even when the prior session's plan.md still had `[ ]` items the user
wanted to keep working on. `working['in_plan_mode']` isn't replayed by
the log restore either, so the card never came back.
Make the baseline conditional: default to 0 so the scanner walks the
full restored history; only push past it when the restored plan.md is
already all-done, suppressing the stale ✓ card. Resolved-path missing
or empty items naturally fall back to "no card".
/stop now also refills the input box with the just-sent user message,
so an interrupted send is one keystroke from edit-and-resend. Only
when the box is empty (don't clobber a half-typed follow-up). Agent
history is untouched — a resend duplicates the turn in LLM context;
/rewind 1 is the manual escape.
Behaviour matrix (/continue):
unfinished plan → baseline=0, scanner finds path, card resumes
completed plan → baseline=len(msgs), card stays hidden
plan.md deleted → resolve_path None, card stays hidden
fresh enter_plan_mode after continue → working flag re-arms, normal
nianyucatfish
7d001ca8a7
