Reading the transcript

The assistant panel draws one conversation the same way, whichever assistant answers: your turns as you typed them, the assistant's words as safe Markdown, and every action as one compact tool line. Claude Code's answers are marked Claude Code, and the in-app assistant's Assistant; the description below names Claude Code, and holds for both.

How the conversation is drawn

Your turns sit on a band marked You; Claude's are headed Claude · unverified — the full the model's words, not a verified result is its tooltip and what a screen reader reads. Claude's words are shown as safe Markdown — headings, lists, quotes, emphasis, strikethrough, task checkboxes (☐/☑), a drawn rule (---), inline code and code blocks (lightly tinted for Java, XML, YAML and JSON) — and everything else as its characters: HTML is shown as text, an image is [image: …] and its address (never fetched), nothing is run. A table renders as an aligned, bordered grid when it fits the panel's width, or as stacked Header: value lines when it does not — never as raw pipes and dashes. Terminal escapes and invisible or direction-reversing characters are shown as marks (␛[31m, ⟨U+202E⟩), so what you see is what you copy. Each answer has Copy (its words as written, never a tool line; it shows under the pointer or with the keyboard focus, and stays shown while the pointer is anywhere over that answer, Copy included) and each code block Copy code; how long the turn took is the answer's tooltip. Each turn shows when it was, quietly, at the right of its header line (14:02; the first turn of a day adds the day, Sun 18 Oct · 09:05); hovering it, or a screen reader, gives the full local date and time, and a walk's recorded dialogue shows recorded and when. A session's first note says what it is and how it is paid for (Claude Code · Safe · uses your Claude plan), once, and again after a restart; a copy that holds such characters says how many.

• Links open only when you click them and confirm a dialog showing the scheme and host on their own line, the host in its plain ASCII (punycode) form — so a look-alike address cannot pass — and only for http and https.
• Chips. A record number (record #12), a node written as code or after the word node, or a chart's name becomes a small button that only navigates — it selects the record, shows the node, or opens the chart; its words stay part of the answer's text, so a selection you copy keeps them. An answer written about an earlier load of the log (before it was reloaded) shows its chips muted, and a click says why and goes nowhere.
• Tool calls are one line each — ▸ analyser_read ✓, refused, denied by you, approved by you · ✓ — and a click on ▸ (a 24-pixel target) shows the call's input. A screen reader hears the line's own words. Alt+↑ / Alt+↓ move between turns. Light and dark themes both work.
• When your full Claude Code is in force, the composer has a coloured border, the Assistant control reads ◆ Full Claude Code ▾ (its tooltip: Full Claude Code (\<mode>) — on: it runs at the next Start · your tools and permissions, then — running) — with ◆◆ and a stronger colour for acceptEdits and auto, which ask about fewer actions. That colour is full Claude Code's alone: a denied call is a neutral ink with its words, a status note its own colour. Send reads ◆ Send (named Send to full Claude Code), the composer Ask full Claude Code…, and the tab and the assistant's own window carry the same ◆ marker. The marker follows the session, so it appears and goes with what the analyser actually started.

Tool lines and their marks

Each action the assistant takes is one line, ▸ and the tool's name, then a mark. Click ▸ to open it: a Claude Code line shows the call's input as it was sent; an in-app line shows the analyser's result, the same result the model was given (up to 4,000 characters on screen).

The in-app assistant's marks:

Mark	Means
✓	the call ran
refused: …	the analyser refused it, and says why
denied: not run (…)	it did not run, for example because the turn was stopped first; the reason in brackets is shown when there is one

Claude Code's marks:

Mark	Means
✓	the call ran
refused	the analyser refused it, and the line says why
approved by you · ✓	you allowed it in the approval dialog, and it ran (approved by you · refused if the analyser then refused it)
allowed: it changes nothing · ✓	a sub-action that changes nothing stored (playing a walk, say), allowed without asking, and it ran
allowed: it creates or shows a chart · ✓	a graph call that creates a chart or only changes what is shown, allowed without asking, and it ran
allowed: writes a new report inside this project · ✓	Full Claude Code only: a report call that writes a new report (or adds to one this session wrote) into the open project, or a new file in its exchange folder, allowed without asking, and it ran
denied by you	you denied it in the approval dialog
denied: not answered within 120 s	the approval dialog timed out
denied: Stop	you pressed Stop while it waited, or as it arrived
denied: another approval was waiting	eight requests were already waiting behind the open dialog, so this one was not queued
denied: too large to show you in full	the request could not be shown in full, so it was not asked
denied: not allowed in Safe analysis	Safe analysis does not offer the tool
denied: not available in this Full Claude Code mode	the mode in force hides the tool
denied: answer in the conversation instead	Claude tried its own question tool; answer in the composer
denied: the mode stays as you chose it	Claude tried to change the permission mode
denied: file exchange is off in the analyser (Settings ▸ Assistant ▸ Allow assistant file exchange), so this write would be refused — the person was not asked	it writes a file and file exchange is off, so you were not asked
denied: the analyser's settings for Claude Code changed	you changed a Claude Code setting while it waited
denied: the request belongs to an earlier turn or session	the request arrived after its turn or session had ended
denied: the turn or session ended before the person answered	the turn or session ended while it waited
denied: this analyser tool is pre-allowed, so an approval request for it is not expected	an approval was asked for a verb that never needs one
denied: the person's own fluxtion-analyser registration is not this analyser	Claude tried the tools of your own fluxtion-analyser registration, which is not bound to this window
denied: the analyser does not recognise this tool	Full Claude Code asked about a tool the analyser does not know
withdrawn by Claude Code	Claude Code withdrew the request before you answered

Approvals explains when a dialog is shown.

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 (Ask the assistant).