A management consultant runs three concurrent client engagements. Each morning, they want to continue exactly where they left off the previous evening on each of the three matters — same context, same tool results, same partial conclusions. Right now they're re-pasting long context summaries at the start of each session, and the consultant is convinced something has been getting lost in translation each day.
What's the right pattern?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Named session resumption (--resume <session-name>) is the documented pattern for continuing prior work. Each matter gets its own named session; resuming brings back the entire conversation, all tool results, all reasoning. The "lost in translation" complaint is the consultant noticing what manual summaries always lose — the architecture of prior reasoning, not just its facts.
File-based summaries preserve facts but discard reasoning. Each morning the new session has to re-derive what yesterday's session already understood, and small details (which option was tentatively rejected, why a particular framing was chosen) silently disappear.
Prompt prefixes don't isolate context — the agent's reasoning about Acme accumulates in the same conversation as its reasoning about Beta and Gamma, and they cross-contaminate. By turn 50, Beta's competitive analysis is influenced by tool results from Acme's, lost-in-the-middle effects make older context unreliable, and "client switching" by prefix is paying probabilistic-compliance costs to fake what session separation gives you for free.
"Re-establishment forces re-grounding" sounds disciplined but is actually a rationalization for not knowing the platform. Session resumption is reliable; the concern is the absence of forced re-grounding, which is solvable by deliberately re-reading the prior conclusions, not by throwing them away.