Diagnose in the analyser¶
The analyser's assistant assembles the selected record(s), the node-type map and the relevant source, then asks a model to explain what happened in the cycle and why.
Its real power is the round trip. The assistant doesn't just answer in a chat window — it can drive the analyser and file its findings back as things you can see and interrogate: a plotted graph, a filtered table, jumped-to and flagged records, all over the same data you can click into and trace to source. The conclusion arrives as a chart, not a wall of text you have to trust — so the verification loop is built in. When the assistant says "the quote calculator stopped because the venue disconnected here", it can plot that series and flag those cycles, and you watch it on screen.
Fault-finding in practice¶
- Scope it. Drag the Time range window to the incident, then select the suspicious records in the table (Shift/⌘-click for several).
- Ask. Right-click ▸ Ask the assistant: What happened here? (or Compare these rows for several), the toolbar's Ask, or Ask about this in the record detail. The question is placed in the composer — see Asking about what is on screen — and goes when you press Send.
- Watch it work. The assistant reads the surrounding records, aggregates over the index to spot the pattern, and when it finds the cause it plots the offending series into a graph (captioned with its reasoning), flags the culprit records, and filters the table down to them — the fault rendered in your own instrument, not described in prose.
- Verify. Click the plotted points, jump to the flagged records, trace a nodeLog line to the exact source method. Nothing is taken on faith; the evidence is right there to challenge.
Which assistant, for what¶
The assistant panel is one panel, whichever assistant answers in it. Set up AI… chooses how it works on this machine, and the mode chip in its header (Safe ▾, ◆ Full ▾, In-app ▾, Off ▾) switches it (Assistant profiles):
| You want | Choose | What it means |
|---|---|---|
| Answers about evidence, including logs you did not write | Safe analysis (the default) | Your installed Claude Code reads your evidence: Read, Grep and Glob in the project folder, and the analyser's own views. No shell, no edits, no web. A call that changes what is open or writes a file is asked of you first. |
| Claude Code that can also edit, build and run in your project | Full Claude Code | Your own Claude Code, with your settings, skills, hooks, MCP servers and tools, under your own permission rules. The analyser does not confine it to the folder. |
| An assistant without Claude Code, on an API key | In-app assistant | The analyser's own assistant, with a provider, model and key you choose. It cannot read your project folder or use your own tools. |
| No assistant at all | Off | Nothing is sent from the analyser. Every view, chart, report and walk still works. |
No profile is local-only. What you ask, and the evidence the assistant reads, is sent to its model provider: Anthropic for Claude Code (on your Claude plan or an API key), or the provider you chose for the in-app assistant. What it costs: for Claude Code on a Claude plan, usage counts against your plan and no dollar figure is shown; for Claude Code on an API key, each finished answer shows its cost, and Settings ▸ Assistant ▸ Limit spend per session (off by default) can cap a session. The in-app assistant shows no cost: your provider bills your key.
Prefer to work from your own terminal or IDE client? That is Diagnose an incident: the same verbs, driven from outside.
In this section¶
- The assistant panel: the header, the composer, the in-app assistant and Copy prompt.
- Ask the assistant: every question you can ask from the screen, and where it is.
- Claude Code in the analyser: turning it on, Start, Safe analysis and Full Claude Code, using it.
- Reading the transcript: chips, tool lines and their marks.
- Approvals: what the approval dialog says, and what Allow once lets through.
The work mode (AI ▸ Work mode ▸ Investigate / Build) tells an assistant whether this session is for investigating or building; Runbooks, glossary and posture explains it.
The canvas¶
Every assistant renders onto the same views you use yourself:
- Records, detail & filtering
- Graphs
- Topology & step-through
- Source navigation
- Investigation reports
- The Project panel
Related pages in other sections: The start page and the sample prompt pack under Get started; Projects, profiles and sharing and Log-source plugins under Administer and secure.
Where the old Analyser assistant page went¶
The single Analyser assistant page was split by task in October 2026. A link to one of its sections lands here; this is where each section is now.
| Section | Now |
|---|---|
| One assistant, two modes; The in-app assistant; No API key? Copy-prompt mode | The assistant panel |
| Choose how the assistant works; Setup | Assistant profiles |
| Asking about what is on screen | Ask the assistant |
| Claude Code in the Assistant tab | Claude Code in the analyser |
| How the conversation is drawn | Reading the transcript |
| Approvals in the analyser | Approvals |
| For administrators | Managed policy (planned) |
| Actions the assistant can take; Connect an MCP client | Verbs |
| Is the log whole? | Log format ▸ Is the log whole? |
| Ask it to show you; Save it as a walk; Walks with a conversation; Send it as a reel | Spotlights, walks and reels |
| Set up a project with Claude | Set up a project with Claude |
| Reading coverage results | Topology ▸ Reading coverage results |
| Your first session; Hand the diagnosis on; Does my client launch the analyser?; If it isn't working | Diagnose an incident |
| Saved charts before opening a log; Explicit session restoration | The Project panel, Working across projects ▸ Restore an earlier session |