You are reading the built-in manual

These are the pages AURA carries within itself: inside the application, the ? key opens the one for the screen you are on. They are published here exactly as they are, so they can be read without installing anything.

The editor for ~/.claude/settings.json, the file that governs how Claude Code behaves on this machine. Four tabs: General, Permissions, Interface & updates, and JSON.

The form and the JSON are the same file

There is only one source of truth: the text of the file, the one the JSON tab shows. The form tabs are only views onto that text — each control shows, next to its label, the exact key it drives, so the mapping never has to be guessed.

Practical consequences:

• a key I do not know stays intact. The screen rewrites only the key you touched; it never erases what it cannot display;
• the order of your keys is preserved — new ones are appended at the end;
• on the other hand, the first change made through a control renormalises the file's formatting to two-space indentation. It shows up in the diff: it is the only thing that changes beyond your own edit.

The Format button in the JSON tab applies the same normalisation, changing nothing else.

As long as the text is not valid JSON, a banner says so, the forms are greyed out and saving is blocked: it has to be fixed in the JSON tab. The parser's error message is shown there as it is, position included.

Inherited, yes, no

Many options are three-position switches, and that is deliberate: an absent key is not false, it means “Claude Code decides”.

• Inherited — the key is removed from the file. The tooltip recalls Claude Code's default for that key.
• Yes / No — the key is written explicitly.

This differs from clearing a text field, which writes an empty value into it: a cleared language gives "language": "", not the absence of the key.

General tab

Setting	Key	Note
Language	language	Preferred answer language
Default model	model	Empty = automatic selection
Effort level	effortLevel	low, medium, high, xhigh
Effort per model	modelSettings	Read-only — a table, not a field
Fast mode	fastMode	Three positions
Extended thinking	alwaysThinkingEnabled	Three positions
Editor	editorMode	normal or vim
Word keys	keybindingFlavor	classic or readline
Conversation cache	promptCacheTtl	5m or 1h
Sub-agent cache	subagentPromptCacheTtl	5m or 1h

Effort per model shows only when the key exists, and read-only: modelSettings is a table keyed by model names, not a field. Saying nothing about it would suggest the only effort in force is the global one just above.

Both prompt caches carry the same third position as the switches: Inherited deletes the key, and Claude Code then decides alone — one hour on a subscription, five minutes on an API key, five minutes for everything outside the main thread. A one-hour cache write is billed at a higher rate than a five-minute one; in exchange, the cache stays warm across longer breaks.

Permissions tab

The default mode (permissions.defaultMode) decides what triggers a permission request. Six values, described in the selector:

• default — asks on the first use of a tool;
• acceptEdits — accepts file edits without asking;
• plan — read-only;
• auto — approval with guardrails;
• dontAsk — denies anything an allow rule does not authorise;
• bypassPermissions — skips every request. I show a red banner while that mode is active: keep it for isolated environments, a container or a VM.

Then come four lists of patterns, each with its counter:

• Allowed (permissions.allow) — e.g. Bash(npm run *);
• Denied (permissions.deny) — e.g. Read(.env);
• To confirm (permissions.ask) — e.g. WebFetch(domain:*);
• Additional folders (permissions.additionalDirectories) — extends the working area beyond the current folder.

In Claude Code, deny wins over allow. An exact duplicate is not added twice: typing a rule that is already there simply clears the field.

Interface & updates tab

tui (fullscreen / inline), autoUpdatesChannel (stable / latest), and cleanupPeriodDays — the number of days before Claude Code cleans up sessions and artefacts.

desktopSessionCleanupPeriodDays is its counterpart for sessions opened from Claude Desktop, which the cleanup above spares. It is a ceiling, and 0 — the default — means there is none: those transcripts are kept until something else deletes them.

spellcheck.enabled underlines misspelled words in the prompt. The underlining needs an aspell, hunspell or ispell installed on the machine; without one, the key does nothing. The rest of the block — the checker, the dictionary, the colour — goes through the JSON tab. Putting this setting back on Inherited removes the key, and the spellcheck block if nothing is left in it: a leftover "spellcheck": {} would read as a setting someone put there.

Then a series of three-position switches, whose “Inherited” tooltip recalls Claude Code's default each time: autoCompactEnabled, autoMemoryEnabled, fileCheckpointingEnabled (the /rewind snapshots), spinnerTipsEnabled, prefersReducedMotion, agentPushNotifEnabled, skipAutoPermissionPrompt, remoteControlAtStartup, includeCoAuthoredBy, autoContinueAtUsageLimit, syncClaudeAiSkills, syncClaudeAiPlugins.

The last two govern what claude.ai drops into ~/.claude/skills/synced and ~/.claude/plugins/synced. One quirk worth knowing: only No acts. Syncing is turned on server-side for your account, so writing true does not switch it on any earlier — Inherited and Yes say the same thing to the file.

The status line, finally, appears only when it is configured, and read-only: its structure is too free for a field. Fine editing goes through the JSON tab.

Saving

Nothing leaves as you type. The unsaved pill marks the gap with the disk, Preview… opens the diff, and the write happens only on confirmation — with a backup of the previous version, and a refusal if the file changed in the meantime. Reload drops the pending changes and re-reads the disk.

A missing settings.json is not an error: the screen starts on {} and the first save creates the file.

Other screens write the same file

settings.json is also changed, on their own keys, by the Hooks (hooks, disableAllHooks), Plugins (enabledPlugins, extraKnownMarketplaces), MCP (the three …McpjsonServers keys and enableAllProjectMcpServers) and Active sessions (permissions.allow, through “Always allow”) modules. All follow the same contract, and the anti-overwrite guard keeps two open screens from stepping on each other.