FAQ

How big a log can it open?

Large ones. Files above the memory threshold (Settings ▸ Performance & S3, default 500 MB) are memory-mapped rather than loaded into heap, so multi-GB logs work. Set the threshold to 0 to always memory-map.

Where are my settings stored?

Your own settings are in ~/.fluxtion-analyser/config, in cleartext (this is a single-user tool), including the in-app assistant's API key. A project's settings are in its own <project>/.analyser/project.fluxtion-settings. The Fluxtion build key is in ~/.fluxtion/fluxtion.apiKeyFile. What is stored where, what leaves lists every file.

Will a build replace the log I am looking at?

No. An open, person-owned project's settled GraphML or audit-log output raises an Open / Not now banner. Your current log stays open until you click Open; accepting a log uses the offered graph and existing pairing checks. Not now dismisses that revision, not future builds. These explicit local opens work with assistant file exchange off. The sample and evidence-bundle working copies are not watched.

The watcher samples bounded project metadata every four seconds and needs another matching sample before offering. It does not read contents or prove that a paused writer finished; normal open validation still applies. See the security boundary for scope and detection limits.

Is my API key safe? Does it leave my machine?

There are two unrelated keys. The LLM key is stored locally, in cleartext in your config file, and used only to call the provider you configured. The Fluxtion build key, managed from AI ▸ Fluxtion API key…, is written directly to the builder's canonical file and is used only when you regenerate a processor. Neither value is included in a project, an exported settings file or an evidence bundle. The analyser reports build-key presence but never validates or redisplays it. See What is stored where.

Does it work offline?

Yes. Everything except the optional assistant, and the buttons that fetch a template or an example, is local. No assistant profile is local-only — Safe analysis included: what you ask, and the evidence an assistant reads, goes to its model provider. Claude Code in the Analyser assistant tab contacts Anthropic when you press Start (its sign-in check and the bridge connection), before your first Send. With the assistant Off, and External MCP/REST access off, nothing in the analyser goes on the network unless you open from S3 or fetch a template or an example. What leaves the machine has the full list.

What does Claude Code itself send to Anthropic?

Besides your question and the evidence Claude reads, Claude Code has usage data of its own, the same in the analyser's panel as in your terminal. The Where data goes picture in Set up AI does not draw it; it links here. Anthropic's Claude Code data usage page is the authority. In short, as that page stated it in October 2026:

• Metrics: latency, reliability and usage patterns, never your code, prompts or file paths. DISABLE_TELEMETRY=1 turns them off.
• Error reports: Claude Code's own error messages and stack traces, with known secrets, paths and e-mail addresses redacted first; on for Pro and Max sign-ins. DISABLE_ERROR_REPORTING=1 turns them off.
• Feedback you choose to send: /feedback (and /bug, /share) sends a copy of the conversation; a session survey's follow-up uploads the transcript only if you answer Yes.
• CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1 turns all of these off at once. On Amazon Bedrock, Google Cloud and Microsoft Foundry, metrics, error reports and feedback reports are off by default.

The analyser turns none of these on or off. It runs your own installed claude, and the environment variables the analyser itself was started with reach it. Full Claude Code also reads your Claude Code settings, as in your terminal; Safe analysis does not load them. How long Anthropic keeps your conversations, and whether they train models, depends on your account; the same page says so for each plan.

How can I tell whether Claude Code is still working?

The composer shows the last observed step and its elapsed time. After a minute without a stream event, it offers Keep waiting or Stop; elapsed time alone is not proof of life. Keep waiting stays dismissed until a new event followed by another minute of silence. An approval has its own status and the existing modal question, not a stall warning or an unusable Show button. ⋯ ▸ About this session… lists tool-call and turn totals.

Is the assistant's action socket safe to enable?

External MCP/REST access is off by default. When on, it listens on 127.0.0.1 only, needs a per-run token, rejects requests from web pages, and is rate-limited; it runs no shell. The verbs that replace what is open or write files are marked destructive, so your client can ask you first, and file writes need their own switch. The security model is the full statement: what each assistant profile can do, what is asked, the limits that are stated rather than hidden, and what an administrator cannot yet lock. One example of what is not asked: under Full Claude Code, a report call that writes a new report inside the open project needs no approval, while a write over an existing report or file, a delete, and anything outside the project are still asked.

Why does the Graph panel say “Charts not loaded”?

Older versions could save two charts with the same name. The analyser cannot tell which definition a name refers to, so it withholds that chart set and explains the conflict on the Graph panel. Log inspection still works. Both definitions remain saved; the analyser does not rename or discard either.

The message identifies the settings file. Close the analyser before editing global settings, or close the project before editing its profile, so autosave cannot overwrite the correction. Give each chart a unique graph.N.name there, keeping its other settings, then restart or reopen the project. You can also open a different, valid project while leaving the original definitions intact. An ambiguous project or imported chart set is refused before it replaces the current project.

Can I follow a growing log?

Yes: Audit log ▸ Follow (tail) polls an open local file and appends newly completed records live, preserving flags and filters; open charts re-extract as records arrive (a chart showing the whole log grows with it, one at the live edge slides, one zoomed into the middle holds). It works for local text logs loaded into memory: not S3, a memory-mapped file or a binary log.

For an ordinary open, the final EOF record is included even without a closing --- separator, as Format 1 permits. Status, Project and context.log.tailNote disclose that completeness is unknown. Starting Follow reopens that snapshot as a live read (clearing record-bound flags and selection through the normal reload boundary). Its trailing record is then pending until a complete separator arrives. Status and context.log.trailingRecordsPending disclose it; live-read counts exclude it. Pausing the writer does not complete a record. After the separator and its newline arrive, the record appears with all fields written before the separator. Do not add a separator to a live export merely to make it appear sooner.

An assistant can use open {follow: true} after a local log finishes loading, or open {follow: false} to stop, each alone. Unsupported readers refuse; context.log.following and supportsFollow mirror the Follow toolbar and menu state. You can stop Follow with either control.

Why is a value in a record not graphable?

Only top-level numeric or boolean nodeLog keys can be plotted. A number sitting inside a toString() (like price= inside MutableOrder(price=…)) is text within one value, not its own key. See Log format.

How do I connect Codex, Claude Code or another MCP client?

With the analyser open, use AI ▸ Set up AI… and choose Your own AI client connects (it opens there when a client is already set up): choose the client, turn on External MCP/REST access (confirmed first), register the analyser with your client, and its last step says when a client actually reached this window. The same door is the start page's AI line and Settings ▸ Assistant ▸ Set up AI…. The Connect an AI client dialog is behind that screen's More options… (its only door since M77-4), for a project-scope command, removing a registration and Claude Desktop's note. Check connection, on the screen itself, proves the local bridge reaches this live window. The screen can explicitly register Codex or user-scoped Claude Code when their CLIs are available. For Claude Desktop and any other client, Generic MCP setup supplies the exact no-token stdio JSON for you to copy or save to a file you choose; it never guesses or edits a client configuration location. See Connect your client for the full sequence.

How do I check for updates?

Each release is on the releases page; the JBang alias and the download link always point at the latest. The app shows a what's new note the first time you run a new version.

Troubleshooting

Symptom	What it means / fix
A source root shows red	The folder doesn't exist (or moved) on this machine. Fix the path in Settings ▸ Source roots; imported setups keep ~-relative paths but a machine-specific root can still be missing.
Clicking a node line doesn't open source	That node's class isn't under a source root or in a searched Maven repo. Add the root, or add the *-sources.jar repo (Settings ▸ Maven repos). Check the right Event processor is selected.
Rows tinted as parse errors	The record didn't fully parse — but the file still loaded and the raw text is shown. The lenient parser never fails a whole file on one bad record.
Open from S3 fails	The aws CLI isn't installed or the profile/region is wrong. Install the AWS CLI and set AWS profile / region in Settings ▸ Performance & S3.
LLM 401 / auth error	The API key or provider/model is wrong in Settings ▸ LLM. No key? Use Copy prompt instead.
Your MCP client says the analyser is not running	See Troubleshooting the connection.

Keyboard shortcuts

Key	Action
F	Flag / unflag the selected rows
F3 / Shift+F3	Jump to the next / previous anomaly
Alt+Left (or ⌘/Ctrl+[)	Source view: back to the previous file
Ctrl/⌘-click	Source view: navigate to the identifier under the cursor
Enter (search box)	Remember the search term
Alt+1 … Alt+9 (⌘1 … ⌘9 on macOS)	Start page: choose that numbered entry
Alt+R (⌘R on macOS)	Start page: return to the analyser's own list after a prompt pack
Enter / Shift+Enter	Start page box and assistant composer: press the button or Send / start a new line
Alt+↑ / Alt+↓	Assistant transcript: move between turns
← / →	A playing spotlight walk: previous / next step