Web UI

OWL-SDA ships with an opt-in local dashboard that shows a run as it happens: the stage timeline, every agent's messages and tool calls, the shared triple store, the output file, and trend charts for triples, violations, and shapes over time.

The dashboard is read-only. It renders the benchmark output directory from disk, so it never interferes with the run.

Starting it

java -jar target/owlsda.jar --config examples/project-1/config.yml --web-ui

Option	Default	Notes
--web-ui	off	Starts the dashboard on http://localhost:8080.
--web-ui-port	8080	Port to listen on.

The dashboard reads the benchmark output, so benchmark.enabled: true must be set in your config — without it the page shows a "Waiting for benchmark data…" placeholder and OWL-SDA logs a warning at startup. See Configuration for benchmark.enabled, benchmark.output-dir, and benchmark.live-interval-seconds.

The server keeps serving after the run finishes:

Run finished. Web UI still available at http://localhost:8080 -- press Ctrl+C to exit.

Press Ctrl+C when you are done inspecting the results.

The page polls every 2.5 seconds. How often the underlying data changes is set by benchmark.live-interval-seconds (default 15), which controls how often a LIVE snapshot is flushed while a round is still running.

The dashboard

Header and stat row

The header carries the current stage badge (LIVE, GENERATE, FINALIZING, REVIEW_ITERATION_<n>, REVIEW_ACCEPTED, REVIEW_REJECTED), a running/idle indicator, a server connection dot, and the last refresh time. A run counts as "running" when the newest snapshot is younger than three live intervals (at least 30 seconds).

Five stat cards sit above the tabs: Shapes Processed, Violations, Triples, Duration, and Total Tokens (with an input/output split underneath). Duration is measured from the first snapshot of the run, not from the current stage.

The Violations card is clickable — it opens a modal with the full SHACL validation report, backed by /api/violations. Press Escape or click outside to close it.

Stages sidebar

Every snapshot in benchmark-summary.json appears as a row with its shape count, violation count, and duration, newest first. Consecutive LIVE ticks are collapsed into a single row with a multiplier, so only real milestones plus one "still working" entry are shown.

Tabs

Tab	What it shows
Messages	One sub-tab per session role (supervisor, reviewer, worker_0, worker_1, …). Shows that role's full transcript with tool calls and their results merged into single cards, Turtle syntax highlighting, collapsible long bodies, and an auto-scroll toggle. Above the transcript, the role's current named contexts ("Delegation Instructions", "Validation Report", "Ontology Summary", …) are shown as collapsible panels.
Output	The current contents of the file at output-path, with size and Turtle highlighting.
Triple Store	The shared WorkerTripleStore dump (triplestore.ttl) from the latest snapshot.
Trends	Three line charts over the run's elapsed time: Triples in store, SHACL violations, and Shapes processed. Hovering shows a crosshair and a tooltip with the value and how far into the run it was. Unmeasured points (for example violations before the first validation) break the line rather than being plotted as zero.

Live per-role indicators

Each role sub-tab in the Messages tab carries three live figures:

• A busy dot — whether that session is currently mid-prompt.
• The model name, with the provider in its tooltip.
• A context usage percentage, shown only when a context window is configured.

These come from metadata.txt, which SnapshotWriter refreshes on every snapshot:

Field	Meaning
model.<role>	Configured model for the role.
provider.<role>	Effective provider, falling back to client.provider when the role sets none.
busy.<role>	Whether the session is currently waiting on the model.
context.<role>.used	Prompt tokens on the last completed call — not a cumulative total.
context.<role>.limit	The role's context-window-tokens. Zero means unset, and the gauge is hidden.

<role> is supervisor, reviewer, or worker.worker_N. Model and provider are written once per role rather than per worker index, since all pool workers share one configured model. Set client.<role>.context-window-tokens in the config to get a usage gauge at all.

HTTP API

The dashboard is a static page over a small JSON API served by WebUiServer. All endpoints are GET, take no authentication, and are stable enough to script against.

Route	Returns
/	The dashboard itself (index.html, app.js, style.css — a fixed allowlist, no filesystem traversal).
/api/state	{ benchmarkEnabled, benchmarkDirExists, liveIntervalSeconds, metadata, roles }. metadata is metadata.txt as a flat object; roles lists the session roles that have a message log, ordered supervisor, reviewer, then worker_0…
/api/history	The raw benchmark-summary.json array — one entry per snapshot, with stage, timestamp, shapesProcessed, currentViolations, triplestoreSize, durationMs, and token counts. [] before the first snapshot.
/api/messages?role=<role>	That role's message log. TOOL_INVOCATION entries are enriched with a parsed tool: { name, arguments, argumentsAreJson }. 404 with {"error":"unknown role"} for an unknown role.
/api/contexts?role=<role>	[{ name, content }] for the role's current named contexts. Files from earlier rounds (prefixed round_N-) are excluded, so this is the round in progress only.
/api/triplestore	triplestore.ttl as { path, exists, sizeBytes, truncated, content }.
/api/output	The file at output-path, same shape as above.
/api/violations	The most recent Validation Report.txt held by any role (supervisor and reviewer are checked first), same shape as above.

The three file endpoints truncate content at 2,000,000 characters and set truncated: true when they do.

Example:

curl -s http://localhost:8080/api/state | jq '.metadata.stage, .roles'
curl -s http://localhost:8080/api/history | jq '.[-1]'
curl -s http://localhost:8080/api/messages?role=worker_0 | jq '[.[] | select(.tool) | .tool.name]'

Where the data comes from

Everything the dashboard shows is read from <benchmark.output-dir>/live/ — the same directory BenchmarkService overwrites in place as the run progresses. Nothing is cached in the server, so restarting the dashboard against a finished run's directory replays it exactly.

See Benchmarking for the full snapshot layout, the archived runs under archive/<timestamp>/, and the offline plotting script.