Mellos Mapping

English | 简体中文

A live, terminal-native map of bottom-up development for Claude Code and Codex CLI.

While Claude builds your system, a split pane beside the conversation shows the system's layered dependency map: primitive layers at the bottom, dependency edges that may only point downward, ghost nodes for what is designed, a spinner on what is being built right now, and solid green for what is built and verified.

A real session: Claude Code on the left, the map pane on the right. L0 is verified, L1 just lit up, everything above is still a ghost.

Why

Most progress reporting is a task list — a top-down worldview. A Mellos map grows the other way: an upper node can only stand on nodes below it, and the picture makes the discipline visible:

• The ghost design appears before any code. Claude declares the whole intended structure as dashed ghost nodes first; you can veto a bad design while it is still only a picture.

• The spinner is where Claude's attention is. One glance answers "what is it doing right now, and on top of what?"

• Done means verified. A node turns solid green only with evidence (a passing test run). If later work cracks a foundation, the node turns red — a cracked foundation under a spinning upper floor is the most honest status report there is.

• The map is a ledger, not a judge. The tools refuse only structural corruption (an edge pointing upward, a duplicate rank). Workflow is Claude's discipline, defined in the bundled skill; violations are made visible, never silently blocked.

Related MCP server: lxDIG MCP

Install

Two lines inside any Claude Code conversation:

/plugin marketplace add GuangminJu/mellos-mapping
/plugin install mellos-mapping@mellos-mapping

Or one line in a terminal:

claude plugin marketplace add GuangminJu/mellos-mapping && claude plugin install mellos-mapping@mellos-mapping

Requires Node.js 18+ on PATH (Claude Code itself requires Node, so you already have it). No build step: dist/ is committed, so a clone runs as-is — dist/server.mjs (the MCP server), dist/watch.mjs (the pane), dist/mmap.mjs (the mmap toggle), dist/hook-session-start.mjs (the SessionStart hook that hooks/hooks.json registers) and dist/store-paths.mjs (the store's path vocabulary, which the plain-node pane launcher imports instead of restating filenames).

The first session after installing asks you one question — how eager mapping should be — and records the answer for every project you will ever open. From then on the hook carries it into each new session by itself; there is no per-project setup step. See Setup: choose when maps open.

The mmap terminal command (the pane's open/close toggle) installs itself on Windows: the same hook notices on session start when the shim is missing or points at an older install, writes mmap.cmd (cmd, PowerShell) and mmap (git-bash) into %LOCALAPPDATA%\mellos-mapping\bin, appends that one directory to your user PATH, and tells you so through the assistant. The PATH change reaches only new processes — and a new tab of a running Windows Terminal inherits the old environment, so close the terminal app entirely and reopen it before the first mmap. The PATH edit keeps the installer's guarantees: nothing happens when the entry is already there, and a PATH that setx would damage (flattened %VARIABLE% references, truncation past its limit) is refused outright, with the entry to add by hand named instead.

The step behind it is still a command of its own, for the cases the hook does not cover — --uninstall, or re-adding a PATH entry you removed while the shims stayed put:

node "<plugin dir>/scripts/install-mmap-command.mjs" [--uninstall]

(--json prints the install outcome as one JSON line instead of prose — the mode the hook itself calls it in.) npm i -g mellos-mapping provides the same mmap via bin, no shims involved.

Update

claude plugin marketplace update mellos-mapping && claude plugin update mellos-mapping@mellos-mapping

Two steps because plugin update compares against the locally cached marketplace clone — the first command is what actually pulls this repo. Restart Claude Code to apply. Releases are version bumps on master. (In-app, /plugin opens the same management UI.)

Upgrading from 0.19

0.20 moved the store out of .claude/ — the map belongs to this tool, not to one client — into .mellos/. The server and the watcher each perform the move once at startup, and print exactly one line on stderr when they do (mellos-mapping: moved the legacy .claude map store to .mellos/ — commit the move.):

Nothing is merged and nothing is overwritten: a project that already has a .mellos/ store (map, pages or config) is left untouched, whatever the legacy directory still holds. If you keep your maps in git, commit the move — git add -A .claude .mellos records it as renames rather than as a pile of deletions plus untracked files.

That one move is the only time either process touches .claude/. Afterwards the tools write nowhere but .mellos/, and never outside the project directory they resolved at startup.

Codex CLI

The same repo doubles as a Codex plugin (codex-cli 0.147+). Three lines:

codex plugin marketplace add GuangminJu/mellos-mapping
codex plugin add mellos-mapping@mellos-mapping
node ~/.codex/plugins/cache/mellos-mapping/mellos-mapping/<version>/scripts/codex-register.mjs

The first two install the skill (the map discipline) as a Codex plugin. The third registers the MCP server at user level — needed because Codex spawns plugin-bundled MCP servers inside the plugin cache with no way to see your workspace, so a bundled server would write the map into the cache. A user-level codex mcp add entry (which the script writes) inherits each session's working directory instead: the state file lands in your project, same as under Claude Code. The registered path is version-specific — re-run the script after updating the plugin.

To watch the live pane beside a Codex session on Windows, run node <plugin root>/scripts/open-pane.mjs <project dir> — it splits the terminal window hosting the session (or falls back to a dedicated "mellos-mapping" window; --window picks that on purpose). Add --page <slug> to open on a particular page — and with a pane already open, rerunning with --page retargets it instead of opening another. The pane auto-follows the page being written — the map the agent is operating on right now; press f to toggle that (a manual page switch also turns it off), or start with --no-follow. Elsewhere run node <plugin root>/dist/watch.mjs from the project directory in a second terminal (or any terminal split). Both take the same flags — see Pane flags.

Any MCP client

The server ships on npm, so any MCP client (Cursor, Windsurf, Zed, Gemini CLI, …) can run it with a standard stdio entry:

npx -y mellos-mapping

The map file lands in the client session's working directory (.mellos/map.json). Open the live pane from the same project:

npx -y -p mellos-mapping mellos-mapping-watch

The server picks its project directory in this order: MELLOS_MAPPING_CWD (an explicit override, for clients that spawn servers from a fixed directory), then CLAUDE_PROJECT_DIR (what Claude Code sets for plugin MCP servers), then the server process's own working directory. Set MELLOS_MAPPING_CWD when your client would otherwise start the server somewhere other than the project you are working in.

The skill/discipline layer is Claude Code + Codex specific; other clients get the five mmap_* tools and the pane, and bring their own prompting.

Use

1. Ask Claude to build something non-trivial. The bundled skill has Claude declare the ghost design and keep the map current as it works.

2. The pane opens itself. Every write tells Claude whether anybody is actually looking (see Who is watching), and Claude opens or retargets the pane with mmap_open when nobody is — you never have to remember to. Open or close it yourself with mmap in any terminal, or /mellos-mapping:mmap in the conversation (Windows Terminal split on Windows, tmux split inside tmux, or a printed command to run in any second terminal). On Windows the pane opens in the terminal window hosting YOUR session, even with several windows open; pass --window to put the map in its own dedicated window instead. Prefer --ascii if your font lacks box-drawing glyphs.

3. Watch nodes light up from the bottom. Interrupt when the picture worries you — that is what it is for.

The pane is mouse-aware (xterm SGR any-event tracking — the same protocol htop and tmux speak):

Every other key is inert, on purpose: an escape sequence the pane does not know (F-keys, Home/End, PgUp/PgDn, Insert/Delete, modified arrows) is consumed whole and does nothing, rather than having its payload bytes read as hotkeys.

Zooming scales the picture first and switches display mode only at the ends of the ladder, so every level still shows meaningful data:

detail+ ← detail ← 100% ← 85% ← 70% ← 55% ← overview

• zoom in past 100% — detail unfolds evidence and the first three lines of the design notes inside the boxes; detail+ widens them into reading cards (up to twelve note rows);

• 85–55% — whitespace tightens and labels truncate proportionally, boxes stay boxes;

• below 55% — labels would stop meaning anything, so the map AGGREGATES: each declared group (a labeled subsystem within a band) becomes one box named foundation subsystem 1/2 with its status derived from the members, edges collapse onto the groups, ungrouped nodes stay themselves. Like a real map, zooming out shows province names — not anonymous dots. (A map with no groups falls back to a pure glyph constellation with per-band counts.) The footer always names the level.

Below the map, between it and the hint line, sits a fixed-height detail panel: a separator you can drag, a status-colored header, the focused node's evidence, both wire directions (uses → … · used by ← …, each neighbour carrying its own status glyph) and its design notes, word-wrapped. With nothing focused it shows the map's dashboard instead. Fixed height — details never float over the map and the layout never jumps.

Glyphs

One status, one glyph, everywhere a map is drawn — the pane's boxes, its tab strip and detail panel, and any other client reading the same store:

The legend under the picture names the four statuses; □ done, no evidence joins it only on a map that actually contains one — the four are the vocabulary, that one is a rule being broken here and now. Documentation diagram kinds replace the status legend with their node-kind glyphs.

Pane flags

Both the pane launcher (scripts/open-pane.mjs <project dir>) and the watcher (dist/watch.mjs) take the same watcher flags; the launcher forwards them verbatim and rejects anything it does not know rather than dropping it.

Launcher-only: --window opens the dedicated "mellos-mapping" window instead of splitting the session's window, and --force opens another pane even though one is already running for this project. Watcher-only: --file <path> names the default page's state file (the launcher derives it from the project directory).

The mmap command

mmap, typed in any terminal, is a toggle: it opens the map pane for the project you are standing in, or closes the one that is already open.

The project is found the way git finds its root: from the current directory upwards, to the nearest one holding a .mellos/ store. Standing in a project that has no map yet is fine — the pane opens on its standby screen and says so until the first mmap_declare.

Closing goes through the store rather than through a signal: mmap writes a one-shot request beside the map, the pane consumes it on its next poll (250 ms by default) and exits, handing the terminal back exactly as it found it — mouse reporting off, cursor visible. A pane still on the standby screen closes the same way. The request is deleted as it is read, and a leftover from a pane that died is swept when the next one starts, so a stale request can never close a fresh pane.

Every watcher flag above works here too, forwarded verbatim; an unknown one is a usage error, never dropped in silence. mmap needs installing once unless you have the npm package. Inside a Claude Code conversation, /mellos-mapping:mmap opens the same pane.

Pages

A project can keep several maps side by side — one effort = one page. Claude targets a page by passing page to any mmap_* tool; the pane grows a tab bar as soon as a second page exists. The active tab is bold in its map's aggregate status color. Each page remembers its own pan, zoom and pinned node.

By default the pane follows the page being written — the map the agent is operating on right now — so declares and updates bring the audience along by themselves. f toggles that, a manual page switch turns it off, and --no-follow starts it off; with follow off, a background page's change lights its tab in status color instead of stealing your view. An explicit --page outranks follow, and a page requested before it exists is shown the moment it appears.

Deleting a page. An effort ends; its page does not have to stay. In the pane, x — or the × the active tab carries when the mouse is on — asks: the footer says press x again to delete <page>, and a second press within three seconds removes that page's file. Switching page, Esc or simply waiting takes the request back. The × sits on the active tab only, so clicking an inactive one switches to it first and offers its × on the next frame. From a tool call it is mmap_remove {pages: ["slug", …]}, applied after that call's map edits. Either way the file is gone for good — the maps are plain JSON, so committing them is the only undo there is.

State lives in the tool-owned .mellos/ directory at the project root:

The map files are plain JSON, safe to commit if you want the maps' history in git. The other three are runtime chatter between a pane and whoever is talking to it — gitignore focus, quit and viewers/ if you commit the store.

Concurrency, stated plainly. Every save is atomic — written to a private sibling temp file and renamed over the target — so a reader polling the store sees the previous complete map or the new one, never a torn write. There is no lost-update protection: two writers saving the same page race, and the last rename wins, silently discarding what the other computed from an older read. Pages are the isolation unit — two sessions that must not clobber each other belong on two pages, which is also the answer to running several Claude sessions in one project.

Diagram kinds

The default kind, dev, is the living progress ledger described above. The same layered-DAG machinery also draws documentation diagrams: pass kind in mmap_declare and the page renders neutrally — plain solid boxes, no ghosts, no spinners, no progress counts.

Node kinds and edge labels work on dev maps too. State machines are out of scope on purpose: transitions cycle, and edges here only point downward.

Sub-maps

A node can link a child page with submap: <page-slug> — the pane badges it ⊞; double-click dives into the child map, Backspace climbs back out. A map of maps, built entirely from pages: no new storage, no new invariants. Whether a node deserves a sub-map is the AI's judgment call — most don't.

Sub-maps are interior detail, not siblings: a page some other page dives into never occupies a tab. Two refinements keep the tab strip from erasing itself — a page whose own node names itself hides nobody, and a link cycle keeps its tabs unless a page outside the cycle dives in, because a cycle has no outside to climb back to. Inside a dive the tab row becomes a breadcrumb — ⌫ parent map ▸ node — and clicking it (or Backspace) climbs back out. When a hidden sub-map changes in the background, the footer says so.

MCP tools

A batch applies bands → groups → lanes → node updates, and within one node update layer moves the node before its other fields, so a node can move and join a group on its new band in one item.

What the boundary refuses, so the ledger never records something it did not mean:

• an unknown key, naming it — a misspelled evidance is an error, not a silently dropped field, at every nesting depth;

• control characters in text fields — an ESC sequence stored in a label would let a map repaint the terminal of everyone who opens it. detail is the exception: newlines and tabs are how a note is written, everything else (ESC, BEL, lone CR) is still refused;

• an empty string where a field is optional — null is how a field is cleared, never a blank that renders as a box nobody can tell from a real one;

• a node whose submap names the page the call itself targets — a link with no bottom, not a parent link;

• a page deletion aimed at the page the same call targets, or at a slug the project does not have — one call must not edit a map it is deleting, and a name matching no page is a typo far more often than a race; the refusal lists the pages that do exist.

A write that does not land answers save failed, nothing changed (retry): the previous file is intact and calling again is the whole recovery.

Who is watching

A map nobody has on screen is a file, not a map — and nothing in the system used to be able to tell the difference. An assistant would declare a design, light nodes up as it built them, and report all of it into a store you had never opened a pane for.

Every pane now publishes a small report while it runs — .mellos/viewers/, one file per pane, refreshed once a second — and every write and every view ends with what those reports say:

The same reports answer "is a pane already open?" for mmap and for the launcher — a question that used to cost a Windows-only process scan and could not say which page was on screen. A report whose pane stopped refreshing it is ignored after five seconds and deleted after a minute, so a killed pane cannot go on claiming an audience.

Setup: choose when maps open

How eager mapping should be is a habit, not a property of a repository — so it is chosen once, for you, in the first session after you install:

• always — map every structured task: workflows, designs, architecture, technical dependencies. The assistant opens the pane on its own initiative; your recorded answer is its standing consent, so it stops asking.

• complex — the same, but only for medium or complex tasks: several modules, a new subsystem, roughly an hour of work or more.

• on-request — map only when you explicitly ask. In a project with no map, the plugin then says nothing at all — zero noise is the point.

The answer lands in <your home>/.mellos/config.json and reaches every session through the plugin's SessionStart hook, which reads it and hands the assistant the matching instruction before you have typed anything. Nothing has to be set up per project, ever again.

One project can still differ: mmap_setup {policy, scope: "project"} records a policy in that project's .mellos/config.json, and a project policy overrides the user one. /mmap setup re-runs the question for either scope whenever you want to change your mind. The policy guides the assistant; it never blocks the tools, and asking for a map explicitly always works under any policy.

Hosts without hooks (Codex CLI, a bare MCP client) get the question another way: while no policy exists in either scope, every mmap_declare reply carries a note telling the assistant to ask you. That note goes quiet for good — in every project — the moment you have answered anywhere.

Structural invariants enforced by the tools: layers form a total order by rank (an integer in 0..99, 0 = bottom, unique per map); every node lives in exactly one layer; edges point strictly downward — which makes the graph acyclic by construction; nodes may not depend on same-layer siblings (if A needs sibling B, either B is really a lower concept or A and B are one node); a group clusters nodes within one band; and node ids and group ids share one namespace — an id names a node or a group, never both, because both render as boxes and one id must mean one box.

Development

npm install
npm run verify

verify is four steps, in this order: typecheck, test, build (bundles dist/, emits lib/ with declarations, cleaning both first), and check:package — which packs the tarball through the real prepack lifecycle and fails if any exports or bin target is missing from it.

The repo is itself layered bottom-up, and each layer has its spec:

dist/ is committed deliberately: plugin installation clones this repo and runs nothing, so entry points ship bundled. CI diffs the committed dist/ against a fresh build, so a source change that forgets the rebuild fails.

Library

The lower layers are also a library (npm run build emits lib/ with type declarations; npm packs it). Subpath exports mirror the source:

Browser-safe means no Node builtins anywhere in the import closure, gated by a test, so a graphical client (a web panel, an editor view) can parse state files and reuse the exact aggregation, zoom and glyph semantics the terminal pane draws with.

License

MIT