Experiments and runs

​

The experiment page

Overview leads with the question and the generated summary, then the authored Overview document.

​

Runs

The Runs tab is the experiment’s run-lineage directory, not just a list. Each family root renders once, followed by its direct children with a └ marker. Deeper runs stay out of the compact directory and are summarised on their direct child as (N more child runs under this). Families are ordered by newest activity, direct children newest-first. Fifty rows render initially with fifty more revealed at a time — but the total count covers every run in scope, including the summarised descendants. When lineage is malformed, the bad edge is detached without dropping the run from the list, and the warning states only what the snapshot proves: parent no longer available, parent unavailable or outside this project, lineage cycle detected, run refers to itself as its parent. These are explanations, not offers to fix anything.

​

Compare

Appears once at least two runs exist. Two views behind one toggle: Both views stay mounted across switches, so toggling never drops keyboard focus or discards fetched series. Compare receives exact experiment membership only — not the recursive descendants the Runs directory shows — because comparing a run against its own child is rarely the question.

​

Versions

Frozen experiment versions pin a run set. That is what a citation points at.

​

The run page

A run has no generated overview; the page opens on Metrics. Related counts every loaded child, including descendants summarised below a direct child. Ancestors and the run itself are not listed — the header and breadcrumb already establish that context. A deeper branch does not expand into another mini-tree; its direct child carries the (N more child runs under this) label, which keeps three siblings visibly different from a three-deep chain.

​

Steps and diffs

/runs/{id}/steps/{step} opens one step: its metric points, its record of non-numeric values, and any sandbox diff captured at that point. /runs/{id}/diff compares captured working-tree state.

​

Trials

/runs/{id}/trials/{trial} is a first-class entity page with the stored title (editable by members), the trajectory expanded into turn and tool-call spans where the format was recognised, the reward, and the captured files.

​

Sessions

A session page carries the transcript and a digest — a condensed reading of what the session did. Sessions attributed to a run appear on that run’s Sessions tab, which is what keeps “why was this launched” answerable.

​

Refresh behaviour

Running entities poll; settled ones do not. A page does not silently swap out the data you are reading — an obsolete request triggered by a workspace or query change is aborted rather than allowed to answer for inputs you have already left.