Super Sidebar

Super Sidebar opens a side pane with the main agent and every subagent as a tree, each with its live status, and under it the steps Claude has taken on your current request. The repository calls it Claude Super Sidebar, and the author built it for the Code tab of Claude Desktop. It makes no model calls, so the pane costs no tokens.

What it does

• Agent tree. The main agent sits on top, with each subagent, background subagent or teammate under its parent. A card shows the status (running, waiting, completed, failed, stopped), elapsed seconds, the number of tool calls and what it is doing right now. Expanding a card adds the model, mode, role, a token total and, for the main agent, the current context size.
• Task flow. Every new prompt starts a fresh flow. Each reply from Claude becomes a numbered stage, titled from the first clause of its text or from the tools it used, such as “Edit file: pane.tsx”. A stage lists three operations with their status; expanded, up to 40. The flow keeps the last 30 stages.
• Plans and verification. A plan submitted from plan mode gets its own card. When Claude reports in a short sentence that tests or a build passed, failed or were skipped, the stage shows it marked as assistant-reported. A completed stage only means its tools ran, not that the code works.
• Preferences. Chinese or English, and white, beige or black cards. The pane refreshes every two seconds and has a Refresh button.

The task flow follows the main agent only. Subagents appear in the tree, with their current tool and call count, but without steps of their own; to read a subagent’s whole conversation, agentpane covers that.

Install

Needs Claude Code 2.1.287 or later. The optional Codex panel also needs Node.js 22.18 or later and a clone of the repository.

claude plugin marketplace add Zafer-Liu/claude-super-sidebar
claude plugin install super-sidebar@super-sidebar

Start a new session and the pane opens by itself; /super-sidebar reopens it. It starts in Chinese, titled 状态: press EN under Preferences to switch that session to English.

How it works

The module watches turn.start, turn.step, turn.complete, agent.spawn and tool.call, and rebuilds its state from the engine’s agent list, the session transcript and the session’s usage. It rereads the transcript every two seconds and after each event, and it follows the text of the main agent’s reply as it streams in. On each tool call it reads the tool name, the description the call carries and the file name of a path argument, to title the step; the code never shows commands or tool results as titles. It does not change or block any tool call, and if the pane fails, the tool result goes through untouched.

State lives in memory for the session. Your language and color choice is saved in the engine’s plugin store, keyed by session ID. The module also reads its own plugin.json and hook files to print the version and a revision hash under Preferences. It makes no network requests, runs no processes and calls no model.

The Codex panel is a separate Node server in the repository, not part of the mod. From a clone:

npm run codex -- --cwd "/absolute/path/to/your/project"

It serves a read-only page at http://127.0.0.1:4317 (--port changes it) that reads Codex’s local session logs, ~/.codex/sessions/rollout-*.jsonl by default, every two seconds. The page shows one conversation and its descendants: the agents, the last 30 steps of the latest turn, the submitted plan and the token usage. It binds to the thread in CODEX_THREAD_ID when started from Codex, to --thread or ?thread= when given, and otherwise to the most recent conversation. It needs no npm install, API key or Codex configuration, and it never drives Codex. It has no login, so it is meant for your own machine.

Limitations

• The author documents the pane for Claude Desktop. The SVG connectors only draw there; in a terminal session the pane falls back to text arrows, and the author says it does not reproduce the visual pane.
• Version 0.2.0 was developed against Claude Code 2.1.299. The author’s own audit says the current code’s install, Desktop loading and auto-open were not rechecked in a live host, and that plan extraction is unverified in the live pane. The portable tests passed before the latest fixes.
• Subagent context size shows as unavailable. Token totals count completed turns, cache tokens included.
• Steps reflect what the transcript shows, not a dependency graph. There is no history across sessions.
• The language choice is saved per session, so every new session opens in Chinese.
• The Codex panel reads a log format that is not a public API. The author checked it against Codex CLI 0.162.1, and it skips ephemeral, remote and archived sessions.