A Claude Code Session Drawn as a Map, With Worn Spokes Where the Work Went

A coding agent's own output is a transcript. It scrolls, it is linear, and by the fortieth tool call you have lost the shape of the session. You know it read six files and ran a test. You do not know whether most of the work went into the shell or into grep, whether the memory folder got touched at all, or how much of the context window is gone.
lot-map is a Claude Code plugin that answers those questions with a pane beside the chat. Every tool call, model step, hook and skill sends a pulse from the centre of the pane out to a station. When something starts, a pulse flies out. When it ends, it comes back. Refusals and errors turn red.
It is a small idea executed with an unusual amount of care about what to draw and what to leave out.
The stations
Eleven stations, each a category of work:
You, for your messages, files sent to you, the task list, the plan and anything you are asked to approve. Model, for each model request and context compaction. Memory, for Claude Code's memory folder plus anything you name in options. Files, for Read, Write, Edit, Grep, Glob, language server code intelligence, and shell commands that only touch files. Shell, for every other command. Network, for web fetch and search, MCP servers, work kept on claude.ai, and shell commands known to reach the network. Browser, for browser tools. Agents, for subagents, workflows, schedules and waiting on background work. Skills, for skills, tool search, and the skill, plugin and connector catalogue. Hooks, for settings hooks that refuse a call, ask permission, add context or take a moment.
The classification of shell commands is where most of the engineering sits, and it is more thoughtful than it needs to be. A program run by its full path is read by its name, including from a quoted path with a space, so CURL.EXE counts as curl. A command behind sudo, env, timeout, nohup or xargs is read as the command it runs. find -exec is read as the thing it is told to execute. So is a command inside for, if, while, ! or a brace group, or a line handed to bash -c or eval. npx counts as Network only with -y, on the reasoning that without it you are mostly running something already installed.
On Windows the PowerShell tool is read the same way with PowerShell's own names. Get-Content and Select-String light Files. Invoke-WebRequest lights Network.
Reading the map
The route is the last few stations the main loop passed through, joined in order, with the newest leg drawing itself. A leg stays for 40 seconds and fades over 20, so a minute after work stops only the spokes remain.
Spokes grow thicker and brighter with use. That is the point of the whole thing: at the end of a session, the map shows where the work went. A session that looks busy but has one thin spoke on Files and a fat one on Shell is telling you something.
Two clock hands grow around a station while a call runs, one turning once a minute, one once every ten. Work sent to the background, a monitor, a background subagent, a workflow, stays lit after the turn that started it ends, with the hands still growing, until the task reports completion. A task that failed leaves the red dot.
Calls that start together fly side by side and the station carries a count, like ×3.
Memory has its own colour. A solid pulse is a memory Claude asked for. A hollow one on a dashed line is what a hook brought unasked. The number under the station counts the first, ↓ counts the second. If you have ever wondered whether your memory folder is doing anything, this is a direct read.
The ring around the centre fills as the session goes, with the percentage above the name. Amber at 70 percent, red at 90. A model change makes the centre and the Model station ring for a few seconds, and the feed gets a row like opus 5.5 → fable 5.2. That also fires when a session starts on a different model than the last session in the same folder, which is a subtle way to catch yourself drifting.
The privacy line, and what it concedes
The README's sentence deserves the full quote: it runs on your machine and sends nothing anywhere, and then, local does not mean blind.
The follow-up explains why. Like any mod that follows tool calls, it is handed each call whole, and it takes names from that and nothing else. There is a PRIVACY.md for what it reads and keeps, and a PRINCIPLES.md for the checks that hold it to that.
I like this framing because it refuses a claim the category makes too easily. A local tool still sees everything you give it. The honest version of "private" is a statement about what leaves the machine plus a statement about what gets read, and this one makes both.
It also means the map is a behaviour summary, not an audit log. A script that opens a connection inside python -c shows as Shell, because classification goes by command name. The text of a here document is not read, so a file written through one shows as Files whatever the text says. A line is read up to its first 8,000 characters. Those are reasonable limits for a visualization and disqualifying for a compliance record. Know which one you are looking at.
Installing
From a Claude Code prompt:
/plugin install lot-map --marketplace alexeyw/lot-map
Answer y to add the marketplace, then choose a scope. From a shell:
claude plugin install lot-map --marketplace alexeyw/lot-map
If it says the plugin is not found in any marketplace, add the marketplace first:
claude plugin marketplace add alexeyw/lot-map
claude plugin install lot-map@lot-map
In the desktop app, click the plus next to the prompt box, choose Plugins, add the marketplace with alexeyw/lot-map, then add the plugin.
The pane opens by itself when a session starts. In a terminal it needs at least 144 columns; below that, type /lot-map and it draws text only unless you ask for the map as a picture or in braille dots. Updating needs two commands, and the first is not optional: claude plugin marketplace update lot-map then claude plugin update lot-map@lot-map, then start a new session. Without the first, the second reports the installed version as the latest.
Built and tested on Claude Code 2.1.321 on macOS, with users reporting it works on Windows.
Why a map and not a table
A table would carry more information. You could sort by station, sum the durations, export the counts. It would also be a dashboard, and dashboards get opened once.
The map works because it is ambient. It sits beside the chat, it animates when something happens, and it accumulates a shape you read without reading. The worn spokes after an hour are a memory aid for a thing that is otherwise invisible: what a session actually cost in attention and context, station by station.
That is the specific problem it solves. Context management is the skill in agent work that nobody can see themselves doing. Having a ring that turns amber at 70 percent, and a spoke that tells you the shell ate the session, makes the invisible part slightly less invisible.
The plugin is at alexeyw/lot-map. Read PRIVACY.md before you install it, which is good advice for the category and unusually easy to follow here, since the author put the concessions in the same paragraph as the claim.
Related articles
Google replaced the model behind Search, and the number that moved was a shell benchmark
Gemini 4 Hydrogen became the default on 12 November, taking over from Gemini 3.8 Flash across Search, the Gemini app and the developer API. It runs about 2.5 times faster, scores 57% on Terminal-Bench 4.1 where its predecessor scored 18.2%, and costs $1.50 in and $7.50 out per million tokens.
Three ways a coding agent wastes money, measured on 30 repositories
TokenEconomizer watches what an agent did in a repository you work in often, then stops it from rediscovering the same build command, loading tools it never uses and looping on one error. Across five agent and model combinations it cut mean cost by 18.4% to 30.9%, cheaper in all 30 repository cells, with Pass@1 up in four of five.
earshot gives coding agents 922 sounds, and every one is rendered from code
An agent can write a game, an interface or a film and never hear any of it. earshot answers requests like "three variations of a footstep on gravel" from a catalogue of 922 designs built from public-domain recordings, with no audio files in the repository at all. It runs on CPU, as a CLI, a Python package, an MCP server and a Claude Code plugin.
The agents stealing API keys were not sophisticated. The gateways were open.
Amazon's decoy servers spent October posing as inference APIs and logged agents built on ARTEX pulling provider keys out of LiteLLM, Dify, Langflow and Open WebUI. The way in was a default master key and a half-finished install. ARTEX itself ships with budgets, scope gates and evidence for every finding.