Activity Map¶

A live and replayable map of what your coding agent is doing. The agent sits at the centre with its context window as a ring around it. Every tool call pulses out to one of eleven stations and back, so you see the files, shell, network, memory and subagents at work instead of reading a list of commands. Afterwards, the map shows where the session's time, faults and context went.

It works the same for Claude Code, Codex and OpenCode, from their session logs. That means you can also replay any past session and scrub through it.

Stations¶

Shell commands are classified by the command's name, not by what a script does internally: python -c "…" that opens a connection shows as Shell.

Reading the map¶

• Pulses. A pulse flies out when a call starts and returns when it ends. Calls that run together fly side by side, and the station shows how many as ×3.
• Worn spokes. A spoke grows thicker the more its station is used, so by the end of a session the map shows where the work went.
• Clock hands. While a call runs, two hands grow around its station: the inner one laps in a minute, the outer in ten. The Model station gets them while the model is thinking with no call open.
• Faults. Errors and denials leave a red count on the station until your next prompt.
• The route. A line joins the last stations the main loop visited, in order; the newest leg draws itself.
• Context. The ring around the core fills as the session goes. It turns amber at 70% and red at 90%, and empties at a compaction.
• Subagents still at work orbit the Agents station.
• Patterns are drawn on the map: a retry streak (a station failed three or more times in a row) glows on that station, and an alternation (three or more back-and-forth cycles between two stations, such as an edit ↔ test loop) is drawn as a double arc with its count.
• The answer. One wide ring leaves the core when a turn finishes.

Below the map, a feed lists the latest activity. Finished calls in a row at one station fold into one row, marked +2 for the two it stands for besides the one named.

Insights¶

For a whole session, the map also reports:

• Time per station: wall-clock time with a call open, and its share of the session's busy time.
• Errors, denials and cancellations per station.
• Context weight: how much tool output each station fed into the context window (characters, and an estimate in tokens). This is where a session's context went.
• Patterns: retry streaks (open, recovered or abandoned) and alternations.
• Models, the slowest calls, hook runs and time, memory pushes and compactions.

In VS Code¶

Open Activity Map in the Agent Hub sidebar, next to the Mind Map. It follows the current session live, including the history read before VS Code started.

• Open Activity Map (the ⧉ button on the view) opens the map large in an editor tab, with insights beside it and a replay bar: play and pause, 1×/10×/60× speed, a scrubber with markers for prompts, faults, compactions and model switches, and Live to return to the present. Keys: Space play/pause, ←/→ ±5 s (Shift ±60 s), Home/End, L live, 1/2/3 speed. With no live session, it offers a recent session to replay.
• Hover a station for its calls, time, errors and top labels; click it to filter the feed and highlight it on the map.
• Export Activity Map… saves the map as SVG or the session's activity as JSON.
• The HTML session report includes the map and the station table.

Settings (sidekick.activityMap.*):

Classification settings apply to new activity; reopen the session to reclassify what was already read.

In the terminal¶

sidekick map --follow draws the live map in any terminal, for any provider. Run it in a split beside your agent:

# tmux: a pane to the right
tmux split-window -h 'sidekick map --follow'

# VS Code: Terminal → Split Terminal, then
sidekick map --follow

It waits for a session if none has started yet and follows newer sessions as they begin (pass --session <id> to stay on one). Keys: f feed, i insights, p or Space pause, ←/→ pause and step 10 s (Shift for a minute), l live, q quit.

Export a session instead:

sidekick map --format text                 # the map, feed and insights in the terminal
sidekick map --format svg -o session.svg   # a still SVG (add --motion full to animate)
sidekick map --format json --at 50%        # snapshot and insights halfway through, as JSON

--at picks the moment:

The dashboard has the same map as panel 9, with an Insights tab.

For agents (MCP)¶

Sidekick's MCP facts server offers get_activity_map: the active session's snapshot and insights as structured data plus a short text summary, optionally with the SVG. An agent can use it to see where its own session's time went, or to notice that it is stuck in a retry loop.

Privacy¶

Activity records hold names, never content:

• a file: the last two parts of its path;
• a web address: the host alone, never a login, password, path or query;
• a shell command: the command's name (and a known subcommand such as git push), never the command line. The descriptive label setting adds the short description the agent itself wrote;
• an MCP call: the server and tool name;
• a hook: its event and matcher, never its command, output or the context it added;
• a subagent: its type.

Control and bidirectional-text characters are removed from every label. Exports, the HTML report and the MCP tool use names-only labels unless you choose descriptive (VS Code setting or --labels).

Accessibility¶

The map has a text description that updates as the session goes, and every station, pattern and replay control can be reached from the keyboard through the insights table and the replay bar. Motion follows your reduce-motion settings.