The question: why do the numbers not add up?

A coding agent might produce provider-native transcript entries and also generate live captured model/network events. Those are overlapping observations, not independent work. Adding the two sets would risk counting the same model call or tokens twice. AgentSight therefore chooses a summary source rather than merging numeric fields blindly. In the released LiveView::build_top implementation, a native summary is considered populated when any of view_events, llm_calls, ortotal_tokens is greater than zero. Otherwise a capture summary is used, if present.

This is specifically the live top overview rule, not a universal policy for every report or stored event. A terminal row can still combine a native session identity with current process resource data and eBPF activity.

Three source-backed examples

The table transcribes three deterministic regression fixtures from the v1.0.35 live_top.rs test module. Values are synthetic test inputs, not measured production traffic. Each numeric triple is events / LLM calls / total tokens.

Input countersTop overviewReason
Native 12 events / 4 calls / 100 tokens; capture 5 / 2 / 4012 / 4 / 100Native summary wins as a whole; captured counts are not added.
Native 0 / 0 / 0; capture 5 / 2 / 405 / 2 / 40Capture summary is the fallback when all three native counters are zero.
Native 12 / 4 / 0; capture 5 / 2 / 4012 / 4 / 0One nonzero native counter selects the entire native summary; no field-by-field token fallback.

The third case is the important counterexample: native events and calls exist, but native tokens are zero. The captured 40 tokens are not substituted. This is whole-summary selection, not per-counter coalescing or a maximum across sensors. Conversely, the second case does not prove capture always has complete data; it only shows what this fallback branch selects when native counters are empty.

What still gets enriched from live processes?

Source priority for top-level model totals does not prevent evidence enrichment. First, build_top obtains live process rows and candidate process trees. It matches local provider-native sessions to live roots using process information, open-descriptor paths, and relevant captured file-audit paths. A matched session row can carry native model/tool/token information alongside the live PID, CPU, RSS, process-family count, and command context. Unmatched live process rows can remain visible separately.

When live eBPF audit events exist, apply_live_capture groups observed exec, failed exit, file and network counters by PID and traverses each displayed process family to enrich its row. It also marks eBPF evidence. These activity counters are a different class of evidence from the overview LLM-call and token totals; their presence does not authorize adding duplicate model totals.

The page also reports explicit provenance notes: native sessions provide primary token and tool information, /proc provides CPU/RSS/process families, and eBPF provides live process activity. The eBPF capture still does not imply that complete TLS payload details are available in top; use the bounded record workflow for that question.

Sections can have their own source preference

Even when overview counters come from native sessions, the live top code separately builds detail sections. For the Network section, a nonempty captured network section may replace the native network section. If there is no useful Processes section, the implementation synthesizes one from visible process families. Recent failures continue to come from the native session snapshot. Thus the selected global counter source does not define the provenance of every row or section.

Reproduce the rule without inventing a workload

The most controlled check is to run the tests included in the exact product revision. Their fixtures do not require real provider credentials or live traffic. Use a checkout of the pinned commit, then locate the tests and run the collector tests with the crate/package configuration documented by that checkout.

git checkout c1574366122a90c451bca6a84713e28813ab95e2
rg -n "live_overview_(does_not_add|uses_capture|does_not_mix)" collector/src/view/live_top.rs
rg -n "local_has_counters|overview_summary|apply_live_capture" collector/src/view/live_top.rs
# Compare the three asserted totals with the case matrix above.

For a real session, inspect agentsight top while an agent is active; preserve both the provider-native transcript and any separately recorded capture before reconciling totals. Compare source identity, time window, and aggregation scope first. Do not assume every observed process event maps one-to-one to a provider model call, or treat a displayed zero as proof of no work.

Scope, limitations, and next decision

This analysis is a source-level audit of v1.0.35, with test fixtures rather than a new runtime benchmark. It does not establish field accuracy for every agent provider, capture completeness, an eBPF runtime overhead result, or whether one particular live session experienced duplicate observations. Fields such as native token usage can themselves be unavailable or delayed.

If the question is why transcript-derived token counts disagree across a completed session, see token reconciliation. If a provider-native session cannot be found, start with native session discovery. For repeatable saved-run evidence, consult the canonical AgentSight documentation.

Versioned primary sources