Use the local dashboard

New to this? Start with Understand your game's performance: a plain-language tour of every part of the page.

Availability: implemented in the maintainer's profiler checkout; not yet a published profiler release. Use a checkout whose --help lists dashboard. The earlier published BPF-only revision does not contain this command. Chromium is the initially verified browser; Firefox verification is pending.

Start and connect

From the implemented vapour-profiler directory:

cargo build --locked --workspace
./target/debug/vapour-profiler dashboard

Open the full URL printed in the terminal, including #token=..., in Chromium. The agent binds to 127.0.0.1:6767. The token is removed from browser history immediately and kept only in page memory. Reloading requires the original URL. Use --port for another local port; a conflict never kills its current owner.

No daemon is installed. Opening the viewer starts no probes or process sampling.

Inspect a dump

Click Open dump and select a .vpt or .folded file. Review the stat cards, hover over the synchronized charts, and adjust the time-window sliders. Click a recorded spike to focus its surrounding window. Search the flamegraph by function name, click a frame to zoom, and select Reset zoom to return.

If the trace has asset records, the Asset load cost panel ranks them by how hard they were to load: click a row or bar to zoom the timeline onto that load. A live native capture has none, so it shows "No recorded assets". See Find expensive assets.

Flame width means retained sample counts, not elapsed time or CPU utilization. Flat samples get no invented callers. Flamegraphs and reports use the whole retained capture, even when the timeline is zoomed. Unsupported metrics remain unavailable. Browser main-thread occupancy is labelled as a lower-bound estimate.

Capture a Linux process

./target/debug/vapour-profiler dashboard --pid 12345 --ttl 300

Replace 12345 with your game PID, then click Start native capture. The API can capture only that CLI-authorized target, not arbitrary page-supplied PIDs.

Current collection reads process CPU, RSS and permission-dependent PSS at 1Hz through /proc. Native FPS, BPF stack sampling and deeper probes are not yet connected. Their visualizations require real imported trace data. Automatic browser forwarding and Vapour CLI synchronization remain pending.

Analyze and compare

  • Offline report: deterministic analysis with no key or external request.
  • Analyze with Claude: optional direct request to Anthropic with a bounded summary. Raw traces and keys never go through Vapour servers.
  • Save report: downloads JSON findings and evidence.

Reports rank observed hotspots, flag sustained memory growth and suggest next steps. Causes remain hypotheses; a growing Wasm high-water mark is not proof of a leak. Unsupported provider findings are dropped; failures fall back to offline analysis. Recommendations do not modify or execute game source.

Analyze a baseline, click Use current summary as comparison baseline, then open and analyze the next dump from the same workload/options. Supported FPS, CPU and memory-growth comparisons appear below the report. The baseline and last-20-actions capture log live only in this tab.

Configure Claude

Open Claude key & model, enter the key, and click Save to keyring. The input clears immediately; this flow uses no browser key storage. It requires secret-tool (libsecret tools) and an unlocked Secret Service keyring. Remove key deletes the saved keyring entry.

Alternatively, the agent reads ANTHROPIC_API_KEY from its launching environment. That value takes precedence and must be unset in the shell separately. Never paste literal keys into shared commands or logs. The initial configurable model is claude-sonnet-4-6. Automated tests intercept key saving and do not make a billable provider call; real keyring/provider use requires your environment.

Export and stop

Export .vpt downloads the retained trace; active native exports are partial. End session, Ctrl-C, SIGTERM, target exit or TTL closes the agent and socket. There is no automatic reconnect polling. Default and maximum session TTL: 300s.

Native captures automatically flush to $XDG_DATA_HOME/vapour-profiler/dumps, defaulting to ~/.local/share/vapour-profiler/dumps. Files use mode 0600. Storage is capped at 200MiB; existing dumps are never overwritten or removed. When full, the agent prints a warning and cannot save another automatic dump—export before ending. Imported files are not automatically persisted.

Budgets and verification

The HTML/CSS/vanilla JS assets total approximately 29KB, with no UI framework, CDN or external fonts. Imports: 50MiB, 1MiB per record, 150,000 records. Flamegraph display: 4096 nodes/64 levels. Native capture: at most 300 metric buckets.

Chromium rendering, imports/exports, flamegraphs, offline hotspot/growth reports, token/origin/Host rejection, native metrics, flushing and off/TTL cleanup passed. Overhead benchmarks and Firefox verification remain pending. This standalone local UI is independent of the staging-only remote Vapour player/dashboard.

results matching ""

    No results matching ""