fusionkit: real model fusion behind your coding agent

The trace dashboard

Pass --observe to boot a local dashboard (the scope app) that streams live fusion signals: sessions, environments, per-panel-model trajectories, and the judge's thinking-to-final flow:

fusionkit codex --observe

fusionkit exports the standard OTEL_EXPORTER_OTLP_ENDPOINT base into every spawned process (spans post to /v1/traces, events to /v1/logs), opens the dashboard in your browser, and tears it down on exit.

What gets traced

Fusion traces with OpenTelemetry: every component emits real spans (units of work, on the traces signal) and events (live point-in-time signals, OTel log records with an event_name, on the logs signal) over OTLP/HTTP, correlated by one trace id and propagated with W3C traceparent/baggage. The vocabulary lives in the semantic-conventions registry (spec/fusion-trace/registry.json):

fusion.turn.info (an event with the environment snapshot)
fusion.candidate spans with live fusion.candidate.step events
GenAI chat spans for model calls (gen_ai.* attributes)
fusion.tool.execution and fusion.cost events
the judge's fusion.judge.thinking events under the terminal fusion.judge span, whose end carries the final answer and marks the session succeeded

Because the pipeline is standard OTLP, the same signals can be exported to any OTel backend instead (PostHog distributed tracing, Jaeger, Tempo) by setting OTEL_EXPORTER_OTLP_ENDPOINT — or the signal-specific OTEL_EXPORTER_OTLP_TRACES_ENDPOINT / OTEL_EXPORTER_OTLP_LOGS_ENDPOINT (and OTEL_EXPORTER_OTLP_HEADERS for auth) — yourself; --observe only fills the default when the variables are unset.

Availability

Published @fusionkit/cli tarballs include a prebuilt, self-contained Scope server at scope/server.js, so npm-installed users do not install dashboard dependencies separately. fusionkit --observe starts that bundled server.

In a source checkout, the app lives at apps/scope. fusionkit-dev deliberately ignores any staged release bundle: it hashes the Scope source, rebuilds when that identity changes, and runs the checkout's Next server. A direct source launch without development mode prefers a staged bundle and uses the same source fallback when none exists. Install the app's isolated dependencies before using that fallback:

cd apps/scope
pnpm install
pnpm build

Maintainers can copy the standalone build into packages/cli/scope with node scripts/stage-scope.mjs. That staging step is for package/release validation; normal source development can use the build-on-change fallback.

When portless is enabled, the dashboard uses a stable local HTTPS name. Without portless, FusionKit prints the raw loopback URL. In both cases, the gateway, panel runner, Python synthesizer, and dashboard share the same trace id for a session.

Product telemetry (separate, opt-in)

Tracing above is local-only plumbing. Anonymous product telemetry is a separate, strictly opt-in pipeline (fusionkit telemetry on), off by default, with a published allow-list of fields. See Privacy and fusionkit telemetry status. DO_NOT_TRACK=1 forces it off.

Correlating a run

Use the run id or trace id printed by the gateway to connect terminal output, session state, and dashboard events. Start with the session list when you are debugging after the process exits.

fusionkit sessions
fusionkit sessions show <id>

The session record gives you the tool, panel, turn count, recent activity, and cost. The trace dashboard gives you the timing and event-level view for the same work.

Logs

Every run writes per-process logs (panel servers, synthesizer, dashboard, cursor bridge) to a temporary run directory, and fusionkit prints the path:

fusion: logs in /tmp/fusionkit-fusion-XXXX/logs

On failure, the most diagnostic lines are surfaced inline and the full log path is included so you can dig in or share it.

Observability