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.
Everything I know about a project, gathered on one screen. Several stores share it, and they do not live in the same place — that is the key to understanding what can be changed and what cannot.
| Block | Real location | Regime |
|---|---|---|
.claude resources | <sources>/.claude/ | Read-only |
| Project memory | The CLAUDE.md files and their peers, in the sources | Read-only |
| Repository documents | README, CONTRIBUTING… at the source root | Read-only |
| Included folders | The source folders you asked to see | Read-only |
| Plans | ~/.claude/plans/ | Deletable |
| Sessions | ~/.claude/projects/<slug>/ | Replayable |
The .claude is read-only banner at the top is a promise: I never write into a project's source folder. Including a folder changes nothing there — it is a display preference, stored on my side.
The indicators
Four tiles: the number of .claude resources, the number of recorded sessions, the number of hooks the project declares, and the weight of .claude — the sum of the sizes of the inventoried resources, not the size of the transcripts.
The first tile and the last speak of the same set, and of it alone: neither the memory, nor the repository documents, nor your included folders enter it, since none of them sits inside .claude. The counter at the top of the navigator announces something else: everything the tree shows, plans and included folders included.
The resource navigator
On the left, the inventory; on the right, the content of the selected file. Groups arrive collapsed: an established project lines up enough of them that the bottom of the tree would need scrolling, whereas collapsed they all fit on screen and say at a glance what the project carries.
Resources are filed by category, inferred from the first segment of their path inside .claude:
Agents, Skills, Commands, Rules, Docs, Tools, then Settings (settings.json / settings.local.json), Memory (every CLAUDE.md), and Other for the rest. Only non-empty categories appear.
Two presentation choices are worth pointing out:
- A skill is presented as a folder — its
SKILL.mdat the top, its reference files below, and its subfolders collapsible. That is the real “one skill = one folder + its references” shape, not a flat list of files. - Rules and docs recover their tree. The list is rebuilt as a collapsible tree, each folder showing how many files it holds.
For memory, the order is deliberately reversed: a level's CLAUDE.md is shown before the subfolders that override it — the order in which Claude Code stacks them.
Files whose name suggests a secret (credential, .env, secret, .key, .pem) are excluded from the inventory, even for reading.
The viewer
A .md is shown with its frontmatter card then its rendered body. The card uses the vocabulary of the category: the agents one for an agent, the skills one for a skill, and for a rule the single paths. For other categories, no vocabulary is assumed: keys are shown as they are, rather than judged “ignored” in the name of an invented reference.
Any file that is not Markdown is shown as a code block, highlighted from its extension.
A document's raw HTML is rendered, not displayed. A hand-written README centres its badges in a <div align="center">, folds a section into a <details> and comments out a block it is not ready to publish: showing those tags in full made the file unreadable. The HTML is sanitised before display — the layout gets through, nothing executable does. An image on a relative path cannot be served from the sources: I show its alternative text in a dashed frame rather than a broken-image icon.
Project memory
This sweep leaves .claude — that is where I walk the source tree. So it is bounded three times: heavy or generated folders are skipped (node_modules, .git, dist, .venv, tool caches…), depth is capped at 6 levels, and the walk stops at 4,000 folders. A pathological monorepo gives a partial list, never a page that stops answering. Folder symlinks are not followed.
The Project memory group gathers the instruction files agents stack level by level: CLAUDE.md, but also AGENTS.md, GEMINI.md, .cursorrules, .windsurfrules and .github/copilot-instructions.md. They are looked for at any depth, because that is how they apply: the file closest to the code wins over the one at the root.
A project can carry one of these without a .claude folder: this block is computed independently.
Repository documents
The Repository documents group shows what a repository publishes: README, CONTRIBUTING, CHANGELOG, LICENSE, SECURITY, CODE_OF_CONDUCT, ARCHITECTURE and a few others, with or without an extension. The README comes first — it is the front door, and alphabetical order would bury it behind an ARCHITECTURE.md. The .github/ templates close the list.
These are taken at the root only. A subfolder's README.md does not have the standing of a CLAUDE.md: in a monorepo, the list would never end.
Included folders
A repository often keeps its most useful matter outside .claude: a docs/, a specs/, a folder of reports. Including a folder adds it to the tree, with its documents (.md, .markdown, .txt, .rst) in depth.
I measure, I do not decide. The picker offers the project's tree with, for each folder, the number of documents it carries — but that number does not say which ones serve you: a folder of templates the program consumes looks exactly like a folder of documentation. That is the one thing I cannot settle for you.
You compose a whole selection there, then apply it in one go. Ticking a folder carries its subtree: its subfolders appear ticked but frozen, since including them separately would make two groups for the same files.
The list is a preference, stored with the others in my install folder — nothing is written into your repository. But it has a second role: I re-read it from disk before every read, and I only open what it covers. That is what makes a requested path never sufficient on its own. With nothing included, this door opens nothing.
Removing a folder from the tree does not touch the folder itself.
The plans
Plans are not in the project: they live in ~/.claude/plans. They are attached to this project because one of its transcripts carries their path (planFilePath). So they appear in the navigator but not in the resource count nor in the weight of .claude.
A plan is the only thing deletable from this screen; that deletion is not backed up.
Hooks
The read-only list of the hooks the project declares — read from its settings.json and its settings.local.json, flattened: the event, the matcher if there is one, and the command. An action without a command (a prompt, http… hook) shows its type instead. These are the hooks that will add to yours when Claude Code works here.
Sessions
The table of transcripts, filterable, sortable on every column.
Session — the session's real title: the one you typed, or the one Claude Code generated and keeps current. A session never named falls back on its first message, and the tooltip says so, so the column does not silently mix two kinds of text. Two possible badges: title set by hand, and presence of sub-agents.
Turns — the human's turns: neither the echoes of tool results, nor the harness injections. It is the same figure as the one in the Diagnostic module.
Tokens — input, output, cache read and cache write added together.
Cost — at API list price, as everywhere else in AURA. A ≥ before the amount means a model in that session has no known rate: the total is then a floor, not an estimate.
Duration — the gap between the first and the last timestamp of the transcript.
Then the git branch, the size of the file and its modification date. On a narrow screen, branch and size are the first columns dropped: they are the ones you find again inside the open session.
A dash — in a measurement column is not a zero: it is a session where nothing was measured (never run, or a transcript with no response). When sorting, those rows are grouped at one end rather than mixed with sessions genuinely measured at zero.
Clicking a row opens the session replay.