Efficiency

Efficiency is a ranked queue of the runs quietly costing you time and money — runs doing extra work the task didn’t need, the span adding the most latency to your slowest workflows, and dependency failures grouped by the external service they came from. Every entry links straight to the evidence (the actual traces) and a next action.

What you see

The Complexity, Latency, and Dependency panels all show at once, not as tabs you switch between — every dimension is on the page together.

A ?tab= parameter only scrolls to and briefly highlights one panel for 2.5 seconds; it never hides the others, and an unrecognized value falls back to no focus at all.

On mobile, the three dimensions collapse into a single-open accordion instead, with evidence shown inline inside whichever item is open.

Summary panels

Each of the three summary panels — Complexity, Latency, Dependency — independently falls back to its own honest empty state if that one dimension lacks data, while the others keep showing what they have.

Optimization Queue

A single ranked worklist across all three dimensions: severity dot, title, reason, linked traces and agents, and a primary action link per item.

When nothing exceeds the current threshold, it says so explicitly and names which dimensions were checked rather than showing an empty list.

Evidence drilldowns

Below the queue, all three evidence sections stay visible together regardless of ?tab=: the Runs doing extra work trace table, latency’s largest-contributing-span breakdown, and dependency failures grouped by target.

Complexity evidence also lists repeated tool calls and retry loops: the same tool called again with the same arguments and getting the same result, or three or more back-to-back attempts at a tool where every attempt before the last one failed. A call that returns a different result the second time (polling, or state that changed) is never counted, and a retry that recovered within two attempts is not waste. Each row names the tool, the agent, how many repeats, and links to the trace.

The dollar figure is a lower bound: it adds up the priced model turn that issued each repeat (shared across the tool calls that turn issued) plus the repeat’s own priced tool cost, and it leaves out the extra context each repeat adds to later prompts. Turns with no price are left out of the sum and counted separately. The detection runs on keyed fingerprints of each tool call’s arguments and result, computed at ingest after redaction, so calls whose arguments were redacted are never matched. Fingerprints exist only for tool calls ingested after this shipped; older traces are not backfilled.

Dependency groups are keyed by span type and operation name, not a stable target host — because that field doesn’t exist on a span yet — so any group using that fallback carries an approximate badge and a note that better attribution requires more instrumentation.

Explain and act

An action rail with three Ask Morse prompt starters and cross-links — into Reliability (only when latency is amber/red), Security settings (only when a dependency reading is red), plus always-on links to Traces, Cost, Quality, and Reliability.

How to use it

Start with the Optimization Queue — it’s the single triage surface across all three dimensions, before you read the individual panels.
Use ?tab= links (or the panels’ own “View evidence” actions) to jump to and highlight one dimension without losing the others.
Follow evidence links straight into the underlying traces or agents.
Use the action rail to ask Morse about a dimension or jump into Reliability, Cost, or Quality for the same window.

Before you have data

Every empty state names a specific, honest reason instead of showing a zero or a blank chart: no data yet, baseline forming, stale computation being refreshed, older data removed by retention, or insufficient attribution for evidence that can’t be reliably derived yet.

The repeated-tool-call section reads baseline forming until a tool call carrying a fingerprint has been ingested in the window (tool calls ingested before fingerprints existed stay unmatched), and insufficient attribution when no tool calls were recorded at all.

If the whole workspace has no traces in the window, the page shows one page-level empty state above three “will populate once…” placeholders instead of three near-identical empty panels.

Traces — the evidence every Efficiency entry links back to.
Reliability — where a latency risk becomes a service target.
Ask Morse — where the “Ask Morse” prompt starters send their questions.