FleetView

A dashboard for a Claude Code agent fleet. FleetView reads the session and transcript files Claude Code already writes on your machine, so you can see every live main agent, every subagent it launched, what each one is writing right now, and what the main agent did with the result. GitHub Copilot Chat sessions are picked up the same way on Windows. Codex main agents and their nested subagents are discovered from CODEX_HOME/sessions (or ~/.codex/sessions).

Codex rows show each thread's own recorded turn state, model, token counts, output and answer. Descendants stay grouped under their parent conversation, including when a subagent forks its parent's history. Tool completion does not mark the agent finished, and a resumed turn returns to running. Quiet Codex families use the same two-hour roster retention as other transcript-backed chats. A turn with no terminal event becomes presumed after 45 minutes of silence; that is not proof that its process stopped. FleetView does not infer a paid cache lifetime from Codex token counts. Use Codex itself to chat, stop or resume these threads; FleetView's Claude-specific controls are unavailable on Codex rows.

FleetView starts nothing while you are not looking. It opens a Claude Code session only where you press a button — 💬 Chat, or a 🔄 Reload that brings a chat back with a keeper. It never stops, refuses or resumes a running agent: control those in the session that dispatched them.

What it looks like

All five are rendered from the shipped panels in a headless browser over an invented fleet — no real session, folder, host or name in them.

The Fleet tab

The tab opens by itself when VS Code finishes starting. It is an editor tab, not a sidebar view; FleetView: Show Fleet brings it back.

A session is a group row: a caret that collapses it, a live or idle dot, the session name, the folder it runs in and its provider, and 🏃 N running · 💤 M standby counted over the rows actually shown. A session whose own read failed says so on its header before those counts, because the counts are the thing the failure makes untrue.

Under each session, one entry per agent. Every column sorts and resizes, and the layout survives a reload:

☰ Columns hides or shows any of them, and a hidden column keeps working — a few start hidden. The toolbar carries the four status toggles (🏃 Running · 💤 Standby · ❔ Presumed · ⏹️ Stopped), each showing the count it is hiding, plus Only me (ticked out of the box, so the roster opens on this window's own work), Show empty (sessions with no rows to show), 🔄 Refresh and ☰ Columns. The strip under the toolbar carries two different things and never confuses them. 🚨 is a read that FAILED: it names the file and the reason rather than rendering as a count. ⏳ is a session file the sweep has not finished reading — nothing failed and there is nothing to do, each refresh carries on with it, and the strip clears itself the moment it finishes.

What the buttons do

Select a row and use the buttons beside 🎯, which names the row they act on.

The PROGRESS line, and the check you can install

Every lane FleetView shows is asked to end each reply with one line:

PROGRESS: <n>% — ETA <m>m — <one clause on what is left>

The 📊 Progress column reads that line — the lane's own claim, never a measurement — and it moves without anything being installed on your machine: the percentage and its countdown come from the lane's own stated ETA against the clock, and when a lane states none at all the column falls back to FleetView's own readings (the steps it has taken, the files it has written, how long since anything moved). Asking for the line was measured failing — 2 of 98 replies carried it when the requirement sat at the end of a brief — so a lane that is never made to state anything still gets a moving column, just a less anchored one.

FleetView: 🔒 PROGRESS check (opt-in hook) — the button in the Fleet tab's title bar, or the same command in the Command Palette — installs a small check into your own Claude Code settings for the case where you want the guarantee rather than the request. A reply that does not END with the line is refused the stop once and answered again with it, which is what makes a lane state where it is.

Nothing is written to any machine until you press it.

🔒 Accuracy hooks — the sharper half, and what it costs

The column moves on its own. What sharpens it is a lane revising its estimate as it learns, and a lane only does that when something asks it to: at the moment it stops, at the moment a subagent starts, and at the tool calls that mean the size of the job has changed. FleetView installs that whole set when it starts.

Agent detail — Output · Diff · Harvest

Click a subagent to open its panel.

• Output — tails that agent's own JSONL transcript. The last 256 KB is read up front, then appends stream in live; assistant prose, 🔧 tool calls and ↩︎ tool results are rendered apart from each other and non-JSON lines fall through raw. It follows the tail and pauses the moment you scroll up.
• Diff — git diff of the agent's worktree against merge-base(HEAD, <default branch>), so committed and still-uncommitted work both appear, untracked files included. It re-diffs every 6 seconds while the agent is alive. Hover any changed line and click 💬 to draft a review comment pre-filled with the agent id and file:line, then ⧉ Copy for main agent to hand it over before the work is merged. An agent that never edits files, and a worktree that has already been merged and pruned, each say so rather than showing an empty diff. Rendering is capped at 60 files and 500 lines per file (400 lines for a file the agent created).
• Harvest — what the subagent handed back and its terminal status, the main agent's next turn, and the tool calls it queued off the back of it — reconstructed from the parent session transcript, never the subagent's own.

Experts

An expert is a Claude Code agent definition in your repo, at .claude/agents/sme-<area>.md, written for one area of the codebase. FleetView shows it on the roster (the 🧠 Expert column), and adds four tools to every Claude Code session on this computer, under the name fleetview:

Every recommend answer also lists every idle instance of that expert, whether or not you asked for it — it is a reading, not a refusal, and which lane the work goes to is your call. FleetView registers these tools at activation, in your Claude Code user settings, so a session started after that has them.

📊 Usage and 💰 budgets

📊 Usage (the status-bar 💰 item, or the command) adds up this week's transcripts: today and this week by model, expert type, host, computer, session and agent; the top consumers; what warm copies saved; and the prices used. The first read of a week of transcripts takes about 40 seconds.

• 🔐 Connect GitHub — signs FleetView in to GitHub once per computer. Each computer then writes its own day files to the private repo fleetview-usage and reads every other computer's, so every computer's usage is added up and named. Until you sign in, the status bar reads Usage not syncing — sign in to GitHub; clicking it runs the same sign-in. The repo is created, private, the first time. fleetView.usage.repo points every computer at another repo instead: owner/name (an organisation repo for a team) or a bare name under your account.
• Sources counted — Claude Code (main sessions and subagents), GitHub Copilot (workspace chats and chats from a window with no folder), OpenAI Codex rollouts under ~/.codex/sessions, and direct API calls your own scripts record.
• Direct API calls — a script that calls a model API itself appends one JSON line per call to ~/.fleetview/app-usage/<user@host>/<YYYY-MM-DD>.jsonl (~/.bullseye/app-usage/ is read too). Python can import scripts/fleetview_usage.py and call record(...); any other language writes: {"ts": <epoch ms>, "app": "...", "provider": "...", "model": "...", "inputTokens": 0, "cacheReadTokens": 0, "cacheWriteTokens": 0, "outputTokens": 0, "totalCostUsd": 0.0, "costKnown": false} — costKnown: false counts the call as unpriced, never $0.
• A computer with no VS Code publishes with node vscode-extension/dist/publishUsage.js from a clone. On Windows, scripts/register-usage-publish.ps1 schedules it every 15 minutes with no window; elsewhere add the cron line that script names. It signs in with GH_TOKEN, the GitHub CLI, or the GitHub sign-in Git already keeps, and reads the repo from FLEETVIEW_USAGE_REPO.
• 📄 TOKEN report (or FleetView: 📄 TOKEN report) writes the same figures to a standalone HTML file you can keep or send, and opens it in your browser.
• 💰 Budgets — set a daily token or dollar budget per project in fleetView.budgets. The 💰 status item then reads Budgets OK or names the project that went over, turns red, and notifies you once per crossing per day.

Status bar

When a subagent settles, a toast reports its outcome and duration with View harvest and View output buttons that open the panel on that tab. Toasts are raised only for sessions belonging to this window, so a busy machine does not spam you from unrelated workspaces.

Commands

What it reads

Every one of these is optional: a file or folder that is not there reads as "none", never as an error.

What it needs

Requires VS Code 1.101 or newer. Nothing else is required; each tool below only powers the part named beside it, and a machine without it still shows the whole roster.

What it writes

FleetView writes only inside your own machine's Claude Code and VS Code files:

• A copy of its own MCP server into this extension's VS Code storage folder, and one fleetview entry in your Claude Code user config naming it.
• The PROGRESS check — a copy of it in this extension's VS Code storage folder, and one Stop plus one SubagentStop hook in your Claude Code settings naming it — only when you press 🔒 PROGRESS check, and taken back out by the same command.
• The accuracy hooks — the same script, registered for Stop, SubagentStop, SubagentStart and PostToolBatch when FleetView starts. Nothing else in that settings file is rewritten.
• Its own usage day files under ~/.bullseye/usage/, the identity ledger ~/.bullseye/identity-ledger.ndjson, and the report file under this extension's storage folder when you ask for 📄 TOKEN report.
• Its own state under ~/.claude/fleetview/ (the reload snapshot, the resume briefs, rows you dismissed) and the reload bridge's answers under ~/.bullseye/fleetview/reload-request/.
• The folder-trusted flag for one folder in ~/.claude.json, only when you 💬 Chat into a folder Claude Code has not trusted yet.

Privacy

FleetView reads your machine and nothing else, with one exception you opt into: once you sign in to GitHub, the 📊 Usage day files are exchanged with a private repo in your own account (fleetview-usage, or the one fleetView.usage.repo names) over GitHub's API. Without that sign-in FleetView makes no network calls; the only other socket is a local one on this computer, between a Claude Code session asking its question and the FleetView MCP answering it.