Claude Super Sidebar
One sidebar that shows your agents, their parent-child links, and the conversation's task flow.
- Source repo
- Zafer-Liu/claude-super-sidebar
- Stars
- ★ 155
- Last updated
- 7d ago
- License
- Apache-2.0
- Primary language
- JavaScript
- FA score
- 58/100 · Major gaps
At a glance
- Works with
- Platform-specificCodex · Claude Code (Partial support)
- You'll need
- Typical use
- You run Claude Code with several subagents and need one view of parent-child links and each agent's running, waiting, completed or failed state.
- Main limitation
- The Claude plugin uses an early-access plugin API developed against engine 2.1.299; the host must support hook modules, claude-code imports and native ui.render panes, and availability may vary by distribution and version.
- Source review
- 58/100 · Major gaps
What does this agent do, and when should you use it?
Claude Super Sidebar is a visualization layer for two different agent hosts. For Claude Desktop Code it installs as the plugin super-sidebar, registers lifecycle hooks plus a two-second refresh timer through hooks/hooks.json and hooks/register.tsx, and renders a single-column native pane titled 状态 · Status from hooks/pane.tsx. For standalone Codex CLI/App it ships a read-only local server, codex/server.mjs, that incrementally tails rollout-*.jsonl files under CODEX_HOME/sessions and serves a browser panel on 127.0.0.1:4317. In both paths a local parser such as hooks/steps.ts groups observed operations into work stages without making an extra model call. The repository includes 32 portable tests split into 9 Claude-side and 23 Codex-side cases.
On the Claude side, the manifest loads hooks/register.tsx via hooks/hooks.json, which registers the /super-sidebar command and a refresh timer at session start. Lifecycle hooks observe agent, turn and tool events, session.messages() feeds the task-flow parser, and hooks/model.ts maintains state and the agent tree. The pane then shows the main agent, subagents with directional SVG connectors, expandable details (model, mode, role, token and context usage, elapsed time, tool-call count), stage-grouped work steps, separate verification conclusions, and preferences for language and card colors. On the Codex side, npm run codex starts a local Node server that re-reads session JSONL every two seconds and renders the bound conversation's latest 30 steps, successfully submitted update_plan entries and observed usage; ?thread= rebinds the page to another conversation, always constrained to the project path given at startup.
- You run Claude Code with several subagents and need one view of parent-child links and each agent's running, waiting, completed or failed state.
- You pair with Codex CLI on a local project and want a browser sidebar tracking the current conversation's stages, tool calls and plan.
- A long session drifted and you want to step back through the last 30 operations and their outcomes by stage.
- You want agent visibility without paying for extra model calls or sending transcripts to another service.
- You are building a Claude plugin and want a working reference for hook modules and a native ui.render pane.
How do you install or deploy this agent?
The Claude plugin needs a Claude Desktop Code host that exposes the native hook-module and pane APIs, a matching Claude Code CLI, and a stable local copy of the repository. Install from the GitHub marketplace:
claude plugin marketplace add Zafer-Liu/claude-super-sidebar
claude plugin install super-sidebar@super-sidebarOr install from a local clone, run from the repository root:
claude plugin install super-sidebar --marketplace . --scope userThen confirm the plugin record and enablement:
claude plugin list --json
claude plugin enable super-sidebar@super-sidebarThe Codex dashboard only requires Node.js 22.18+ and locally persisted Codex sessions, with no npm install, API key, or Codex configuration change.
How do you use this agent?
Start a new Code session after installing so the plugin loads; it attempts to open 状态 · Status automatically, and /super-sidebar in the conversation input reopens it. Use 展开 / 收起 to expand or collapse an agent and 刷新 to refresh immediately; the pane also refreshes every two seconds.
Run the standalone Codex panel from the repository:
npm run codex -- --cwd "/absolute/path/to/your/project"Then open http://127.0.0.1:4317. To change port, data directory or thread binding:
node codex/server.mjs --cwd "/path/to/project" --port 4318 --codex-home "/path/to/codex/data"
npm run codex -- --cwd "/absolute/path/to/your/project" --thread "full-thread-id"Validate the manifest and run the portable test suites:
claude plugin validate .
npm run test:claude
npm run test:codexWhat are this agent's strengths and limitations?
- The Claude path is a genuine native pane (ui.render) rather than an external window, while Codex gets its own read-only local browser panel reading its own data source.
- Task flows are derived by local parsing with no additional LLM call, so there is no extra token spend or model latency.
- It separates "operations completed" from verification conclusions and labels verification as assistant-reported, so exit code zero is not mistaken for correct behavior.
- 32 portable tests run without installing packages, and the repo includes an isolation audit report plus captured verification evidence.
- The UI is authored bilingual (Chinese and English) with white, beige and black card colors, and the Codex panel starts without any API key.
- The Claude plugin uses an early-access plugin API developed against engine 2.1.299; the host must support hook modules, claude-code imports and native ui.render panes, and availability may vary by distribution and version.
- Regression cases added after the latest isolation fixes were never executed, and no runnable compatible Claude Code CLI was found, so native validate, plugin test and Desktop loading were not rerun.
- The Codex panel reads a rollout JSONL format that is not a stable public API, verified only against Codex CLI 0.162.1 and one machine's App session records; ephemeral, remote, archived and unpersisted sessions are excluded.
- The Codex panel is a read-only companion browser UI: it does not drive conversations, discovers at most 100 sessions when unbound, and offers no multi-user authentication or remote monitoring.
- Codex "last request input tokens" is not live context occupancy, and abnormal exits can leave the last observed state on screen.
How does this agent compare with similar options?
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Claude Super Sidebar This agent | 58 · Major gaps | — | ★ 155 | 7d ago | JavaScript | — |
| Ordewell | 67 · Some gaps | CLIFree + model costs | ★ 206 | today | TypeScript | Codex · Claude Code · OpenAI API · Claude API |
| AgentBridge | 61 · Some gaps | CLIFree + model costs | ★ 416 | 1mo ago | TypeScript | Codex · Claude Code |
| Codex Dynamic Workflows (Claude Code skill) | 53 · Major gaps | — | ★ 327 | 3mo ago | JavaScript | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
README states the Codex panel is read-only, reads only local JSONL, does not drive conversations, needs no API key, and adds no models or provider config, so the permission surface is relatively restrained (least_privilege=2). However, installation and enablement have no explicit confirmation step and the plugin auto-attempts to open the pane on session startup, so user confirmation is weak (user_confirmation=1). Data flow is described reasonably in the How it works section and Codex configuration (data_flow_transparency=2). Sensitive data handling only mentions stripping sensitive info from bug reports; storage, redaction, or access control for local session records and token usage is not addressed (sensitive_data_handling=1). package.json declares no dependencies and there is no lockfile or dependency audit note, so dependency security cannot be assessed (dependency_security=1). External effects are limited: read-only panel, local ports 4317/4318, no Codex config changes (external_effects=2). Rollback only covers Ctrl+C and plugin enable/disable; no uninstall, cache cleanup, or state recovery guidance (rollback=1). Source attribution is clear: author Zafer_Liu and personal site in README, Apache-2.0 LICENSE (source_attribution=2).
The README is internally consistent and honest about validation status for both Claude and Codex paths, admitting latest isolation regressions were not run and native engine validation is incomplete (self_consistency=2). Dependency availability declares only Node.js >=22.18 with no third-party packages, but the Claude side depends on an early-access plugin API and engine 2.1.299, making availability highly uncertain (dependency_availability=1). Failure messages are limited to malformed lines being skipped with a notice and missing records showing a waiting state; no systematic error messaging or diagnostics (failure_messages=1).
Audience and scenarios clearly separate Claude Desktop Code users from standalone Codex CLI/App users (audience_and_scenarios=2). Capability boundaries are thoroughly documented: does not drive conversations, no multi-user auth, no remote monitoring, no cross-session aggregation, adds no models (capability_boundaries=3). Trigger precision covers the /super-sidebar command, startup auto-open, two-second refresh, and --thread / ?thread= binding rules, but user control over auto-open and refresh is limited (trigger_precision=2). Environment fit is narrow due to the early-access plugin API, specific engine version, and non-stable Codex rollout format (environment_fit=1).
Information architecture is strong: path responsibility table, anchored navigation, bilingual README, clear How it works diagram (information_architecture=3). Install notes give marketplace and local-clone routes plus verification commands, but rely on unverified host APIs and omit uninstall (install_notes=2). Naming stability proactively explains the claude- prefix restriction and plugin ID difference, but product name, plugin name, and pane titles diverge, adding cognitive load (naming_stability=2). Examples and FAQ consist only of illustrative SVG layouts explicitly declared not runtime screenshots, with no FAQ (examples_and_faq=1). Known limitations section is detailed and honest (known_limitations=3). LICENSE is the full Apache-2.0 text (license=3). Version 0.2.0 appears in badge and package.json but there is no CHANGELOG (versioning_changelog=1). Maintenance responsibility points only to the author's personal site and welcomes issues/PRs, with no governance, support channel, or security disclosure process, and the publisher is unverified (maintenance_responsibility=1).
Output usability: the panel shows agent trees, task flow, status, and usage, which is genuinely useful for observing multi-agent execution, but the Claude native pane is currently unvalidated and the Codex panel is read-only without a session picker (output_usability=2). Marginal value comes from organizing scattered session records into a visual task flow with no extra model calls, an incremental improvement rather than irreplaceable (marginal_value=2). Cost-benefit: no npm install, no extra LLM calls, local parsing, low cost; but reliance on early-access APIs and unstable rollout format adds maintenance cost (cost_benefit=2).
README provides test commands and a historical pass count (32 passed) and links to an audit report and verification evidence directory, making claims traceable (claim_traceability=2). However, only README, LICENSE, and package.json were supplied; test files, audit report, and verification evidence are absent from the evidence, so cross-source corroboration is not possible (cross_source_corroboration=1). Fact/inference separation is excellent: it distinguishes historical evidence from currently unverified items, declares preview images as sample data not runtime screenshots, and stresses that exit code 0 does not prove product behavior (fact_inference_separation=3).
- The Claude native plugin depends on early-access hook-module and ui.render APIs and was developed against engine 2.1.299; current native validation, Desktop loading, and auto-open were not rerun, so real-world availability is unknown.
- The Codex panel reads a rollout JSONL format that is not a stable public API; compatibility is limited to tested local versions and may break after Codex upgrades.
- The plugin auto-attempts to open the pane on session startup without explicit user confirmation, and install/enable docs omit uninstall and cache cleanup.
- Storage, redaction, and access control for sensitive local session records and token usage are undocumented; the panel has no multi-user auth and should not be exposed beyond localhost.
- package.json declares no dependencies and there is no lockfile or dependency audit, so dependency security cannot be assessed; the publisher is unverified and maintenance/security disclosure channels are unclear.
- Test pass counts, audit reports, and verification evidence were not included in this evidence set and cannot be independently corroborated.