Scope: source on master, not the stable release
This audit pins source to commit 9d33c5b6e9d8586960b6246d40a8cfa87502340c, merged on 20 November 2026. At research time, the latest stable release remains v1.0.35, published before the Events view change. Do not assume an installed v1.0.35 binary includes this view. Demo snapshots and code fixtures are not benchmark measurements. This page documents the new master-branch implementation, not a new release or verified production incident.
For an earlier layer—converting raw eBPF timestamps and SSE windows— see timestamp semantics. Here the question is how already-normalized snapshot records are transformed into displayed spans.
Four kinds of span, four evidence contracts
eventSpans() assembles ordered intervals from process nodes, LLM audit events, network targets and network audit events, structured Tool calls and native transcript Tool rows. A common x-axis does not imply identical durations or confidence.
- Exec: a known start and optional process end. If the end is absent, the span stays open through the capture watermark, even when status is merely observed. An explicit exit code determines success or failure.
- LLM: captured audit request/response events match using a call identifier even if calls overlap. Native rows without that identity pair a request with a following compatible response from the same process. The display also de-duplicates matching PID/start/end/label combinations.
- Network: a target row becomes an interval between its first and last observation with aggregate counts. A separate network audit event becomes a point. Neither shape automatically measures the duration of one HTTP call.
- Tool: structured calls can have explicit start, end or duration fields. Native transcript audit Tool events without duration become instantaneous points. The latter lack enough evidence to reconstruct actual latency.
Eight source-level interpretation cases
These examples are synthetic inputs and source-derived outputs, not real workload measurements. The upstreamregression tests exercise the important construction cases.
| Observation | Displayed construction | What it supports |
|---|---|---|
| Process starts at 1,000 ms; no exit; capture ends 20,000 ms | Open bar extends to 20,000 ms | Observed start, but no confirmed exit or 19-second successful execution |
| A child command has exit_code 2 | Failed process bar; detail contains exit 2 | The command exited nonzero; total agent outcome is a separate question |
| Captured model requests A and B overlap | Captured call IDs pair each request with its response | Do not pair each request with the very next response globally |
| The same model call is present in captured and native records | PID, start, end and label can collapse duplicate display spans | A heuristic is not a proof of perfect source de-duplication |
| An endpoint has 40 aggregated observations, two errors | Network target spans first to last timestamps and appears successful | Aggregate color is not proof that every request succeeded or that the interval is one request |
| A native transcript Tool row has only one timestamp | Point-like span with start equal to end | The real Tool duration is unknown; the model did not prove instant execution |
| A Tool reports 3,000 ms start and 6,000 ms duration | Reported interval spans to 9,000 ms | This is the Tool-reported interval, unlike a process or network window |
| A zero-duration row appears in a 10-second chart | Minimum 0.4% draw width produces a 40 ms visual bar | Visual width is not source measurement precision |
Why a green network bar can still include errors
The network target is marked failed only when its count is positive and the error count is at least as large as the count. In the regression fixture, an endpoint with 40 observations and two errors is labeled successful; another endpoint with two observations and two errors is failed. These are UI aggregate statuses, not guarantees that every request completed successfully.
Likewise, the bar stretches from the first to the last observed target timestamp. That duration might contain many requests and idle gaps. It is not necessarily one network round trip, one connection, or a contiguous task segment. Seenetwork target semantics for why counts must not be equated with unique HTTP transactions or exfiltration events.
Rendering changes visible width, not the source interval
The Events chart gives tiny spans a minimum drawn width of 0.4% of the displayed range. In a ten-second range, even a zero-duration point can occupy 40 ms of visual width. Packing overlapping bars onto separate rows and maintaining a small display gap also shape the picture. Color follows normalized success, failed, running or unknown status; it is not an independent verification of the event.
The table initially displays at most 300 visible rows, with an option to expand, while kind and failure filters change which records appear. Absence from the current view is not evidence that an action never occurred. Do not sum all bar widths to estimate wall-clock runtime or resource cost.
Verify the findings without collecting user data
Inspect the pinned public source and tests. These commands are read-only; they do not attach probes or expose local sessions.
git clone https://github.com/eunomia-bpf/agentsight.git
cd agentsight
git checkout 9d33c5b6e9d8586960b6246d40a8cfa87502340c
rg -n "eventSpans|llmSpans|networkSpan|toolSpan|MIN_BAR_PERCENT|TABLE_LIMIT" ext/web
sed -n "1,220p" ext/web/utils/eventSpans.test.mjsIf you need evidence of actual command effects, correlate the event with a system trace and theprocess exit code semantics. For missing fields and exported-data limits, review thesnapshot contract. Source timestamps, native-session data, and Linux capture have different observation boundaries; this visualization cannot reconstruct unobserved activity.
Primary sources and limitations
- Span construction, status and model pairing
- Regression fixtures for open processes, overlapping calls, tools and networks
- Chart rendering, minimum bar width, filters and table size
- Analysis-tab integration
- Master commit introducing the Events implementation
- Stable v1.0.35 release at research time
This is a source/test audit of an unreleased master-branch UI implementation. It does not assert precise latency accuracy, real-world capture completeness, a new runtime benchmark, or a future release date. CLI installation and configuration belong to the canonical docs.