Flightdeck Claude Code mod: a Claude Code agent dashboard for your terminal

Flightdeck is a community mod by Stephen Casella (scasella) that works as a Claude Code agent dashboard: one pane that shows what your session and its agents are doing right now. It gathers the model name and effort, a context gauge, session cost, rate-limit windows when your plan reports them, every permission check, subagent cards, and a receipt for the last turn.

The panes are: main (model, effort, mode, request count, context, cost, speed of the last request), architect (consults of an advisor-style agent or server tool, matched by a regular expression), gate (one cell per permission check: allowed, decided by the classifier or you, pending, or denied), agents (cards that become swimlanes past three agents), loops (model loops that match no card), receipt (duration, agents, edits, errors and cost of the turn) and a short log. Panels with nothing to show take no room. The author's README says every number comes from a real session event; the architect "moments" are inferred and are labelled that way.

The recording starts Claude Code with the mod loaded in the small demo project. The terminal is narrower than 144 columns, so we open the pane with /flightdeck. A bordered summary appears above the prompt with idle and $0.00. Then we ask Claude to read README.md and describe the project in one sentence. While the turn runs, the pane switches to working and a gate row appears with 1 allowed · 0 classifier · 0 denied. When the turn ends, it shows the context gauge (4%), the cost, and a receipt line: last turn 47s · 0 agents · 0 edits · 0 errors.

Use any of the methods in the install block; they install straight from the author's own marketplace. Run /reload-plugins and then /flightdeck. The author notes that the installer may say config options are not set; the defaults work.

Everything is kept in $.state atoms, and each hook updates them and passes the event on. A permission check is recorded like this, from hooks/register.tsx:

on('tool.check', async ($, e, next) => {
  const verdict = await next(e)
  if (e.tool_use_id) {
    const check: Check = {
      id: e.tool_use_id,
      tool: e.tool,
      bucket: bucketOf(e.tool),
      verdict: verdict.decision === 'allow' ? 'rule' : verdict.decision,
      detail: shorten(describeInput(e.tool, e.input), 90),
      // ...
    }
    await update($, gate, g => recordCheck(normalizeGate(g), check))
  }
  return verdict
})

The verdict is returned as it came. The pane itself is drawn on the Pane render site. Layout depends on where it sits: a summary of at most 8 rows inline above the prompt, one column when docked beside the transcript, and two columns from 110 columns. Two small client modules, rail.tsx and elapsed.tsx, draw the animated connectors and the running clocks on the surface's own frame clock so the pane does not redraw. Reducers and formatters are pure functions in hooks/core.ts.

Options are in /config: panels (which panels show and in what order), layout (auto, compact, wide or mini), maxCards (1 to 6), motion (while-active or off), palette (theme or pastel), architectPattern, architectLabel, gateLabel, matchDescriptions, moments, openOnStart and statusLine. The /flightdeck command takes open, close, reset and layout auto|compact|wide|mini. Focus the pane with ctrl+x tab; keys 1 to 6 expand a card and f, s, o open the file, shell and other rows of the gate.

We read every file the plugin can execute before running anything: plugin.json, marketplace.json, hooks/hooks.json, hooks/register.tsx, hooks/core.ts, hooks/rail.tsx, hooks/elapsed.tsx, the type declarations and the tests. There is no package.json, install script, binary or bundled code. A search for network, process, eval, dynamic import and hidden Unicode found nothing.

• Network, processes, files, environment: none. The mod makes no requests, runs no commands, reads and writes no files and does not read environment variables or call the model.
• What it observes: every tool call's name and input, every permission verdict, subagent spawns, token usage and final answers, and the first 70 characters of each prompt for the log. It keeps short summaries in session state only, with credentials masked by a regular-expression filter. The filter is best effort, so treat the pane as showing your commands.
• Gating hooks: tool.check, tool.call, agent.spawn, agent.offer, session.append, session.compact and classic.UserPromptSubmit are all pass-through: they await next(e) and return its result. claude plugin validate notes they have no .catch; the two classic and compact hooks also do not run where an organization manages Claude Code.
• Telemetry, daemons, other writes: none; nothing persists after the session.
• License: MIT, with a LICENSE file.

claude plugin validate reported no errors or warnings. The author's 28 tests passed on Claude Code 2.1.303.

Checked with Claude Code 2.1.303 on macOS (validate and tests); the author's README asks for 2.1.287 or later. If the pane does not show up:

• Run /reload-plugins, then /flightdeck. Below 144 columns the pane does not open by itself.
• Look in the transcript for a dim line starting flightdeck:; it names a failed hook or the reason the pane was refused.
• If colors look wrong, set palette to pastel. If counters look stale after an update, run /flightdeck reset.
• On VS Code and mobile the connectors and clocks are static.