Interaction cost (GOMS-KLM)

An optional overlay for UI rounds: while you drive the product, record what a person would have had to do to reach the same state. The round then carries a user-equivalent interaction cost, so a flow can be judged on what it costs to walk, not just on whether it passed.

Where it surfaces: in that round's report, as its own collapsible section — which on the acceptance page opens from the round history, available to the acceptance owner. It is not part of the shared aggregate a non-owner viewer sees, so never promise a reviewer they will see it there.

Skip it whenever it does not apply. A CLI or backend round has no interaction to price. A machine without a UI driver installed (no agent-browser) records no trace. In both cases you publish exactly as usual — no trace file, no cost section, no warning. Never hand-write the numbers to fill the gap.

The seam: you count actions, the platform prices them

You never compute seconds. The driver appends one JSON atom per action to interaction-trace.jsonl in the report directory, carrying raw operator counts; lh acceptance run ingest finds that file and prices it with the platform's pinned timing model (goms-klm@lobe-v1).

That split is the point: every round is priced by one model, and any published number can be recomputed from its trace. A summary you calculated yourself is not comparable with anyone else's.

jsonl

{"schema":"lobehub.agentBrowserKlmTrace@1","type":"action","phase":{"id":"login","label":"Sign in"},"klm":{"category":"action","operators":{"P":1,"K":1}},"durationMs":840}
{"schema":"lobehub.agentBrowserKlmTrace@1","type":"mental_estimate","phase":{"id":"first-view"},"klm":{"category":"mental","operators":{"M":2}},"mentalEstimate":{"score":3,"confidence":0.75,"reason":"Reading the state and deciding the next action"}}

Field	Meaning
`schema`	Must be `lobehub.agentBrowserKlmTrace@1`; a foreign tag is not summed
`phase.id` / `label`	Groups atoms into a step of the journey; `phase.checkItemId` links a check
`klm.operators`	Raw counts — see the operator table below
`klm.category`	`action` \| `mental` \| `blocked` — `blocked` charges nothing
`durationMs`	Agent wall-clock, reported separately from the user-equivalent price

Operators

Operator	Counts	Typical source
`P`	Pointing at a target	`click`, drag
`K`	One keystroke / button press	`click` (the press), `press`
`T_chars`	Characters typed	`fill`, `type`
`H`	Homing between keyboard and pointer	Switching input device
`M`	Mental preparation — estimated by you	Deciding, locating, reading state
`R_ms`	Measured system wait, in ms — not modeled	Navigation / load the user waits out

Two rules keep the number honest:

An action that did not happen costs nothing. A failed, timed-out, or blocked command is recorded with "category":"blocked" and zero operators. Never charge a retry loop as if a person performed it.
M is an estimate, and it is yours. Record it explicitly at the moments a person would actually have to think — first view of a screen, choosing between options, re-orienting after a state change — with a short reason. Do not sprinkle it on every action to inflate the total.

Publishing

Leave interaction-trace.jsonl in the report directory and publish normally:

bash

lh acceptance run ingest "$REPORT_DIR" --source agent-testing --json

Ingest computes interactionCost and stores it on the round. If a trace exists but records nothing priceable — every action blocked, for instance — the round publishes without a cost section rather than claiming a 0s journey. An explicit result.json.interactionCost is left untouched, so a driver that already computed a summary is never overwritten.