Ask the assistant

Every "ask" in the analyser comes from one list of questions, labelled for the assistant in force — Ask Claude: … in the Claude Code mode, Ask the assistant: … in the in-app one:

Question	Where
What happened here?	right-click one record; the toolbar's Ask and the record detail's Ask about this with one record
Compare these rows	right-click two to twenty records (the same buttons with several); Ask Claude about this diff in Diff selected two records
Did this node run, and what does the evidence establish?	right-click a node in the topology, with a record selected
Explain this point	right-click a chart point or note (Ask Claude: explain this point, beside the chart's own Add explanation…, which is your words)
Explain this finding	right-click one flagged record
Summarise this window	right-click the time slider, or a chart's visible window

• The question goes into the composer, not to the model. It is added after whatever you have written there, only the added lines are highlighted, and the line above the box says Draft — not sent (for Claude Code; the in-app assistant's says Press Send to ask the assistant — nothing has been sent). Nothing is sent until you press Send; in the Claude Code mode with no session running, Send starts Claude Code first. Asking the same question again replaces it rather than adding a second copy, and a question drafted for the other assistant is replaced by the next one you ask. If you switch assistants with such a question still in the box, Send says so in the status line and offers one action, Ask again for Claude Code, which places it for Claude Code.
• It names where to look, never what is there. A question names record numbers, the log's file name, node ids, a chart's name and a time — never a value, a note's words or a finding's text, because those can carry text written to steer an AI. The assistant reads the records itself (analyser_read), so in the in-app mode the questions need Settings ▸ Assistant ▸ Let the in-app assistant change the views as it answers; with them off, the items say so and are disabled. A question's draft never carries the raw text of your current selection.
• A chip by the composer warns when a drafted question is about an earlier log — and shows nothing while it is about the log open now. The question's own words do not name a load; the analyser tells the model which log it checked when it sends the message.
• If the log changes before you Send — you reopen or reload it — the chip reads an earlier log · Send will refuse it — ask again, and a question that names its records is refused, "This question names records of a log that has been reloaded since it was drafted (load 3 → 4) — ask again from the records", and stays editable. Deleting the whole question from the box drops that check.
• During a walk with a conversation (a journey), an ask does not go to any model: it shows the next recorded turn by playing the step that reveals it. When there is none, the status line says "The recorded conversation has no further turn — end the walk to ask yourself".

The start page asks too: its numbered entries and its box place a question for the assistant in force (The start page).