Local transcript evidence

Uncollapse the actual run.

Turn pasted agent activity into ordered events. Nothing leaves this browser, and the parser does not invent calls that are not in the transcript.

Agent-run transcript0 characters

Expand tool evidence Group by event type

Boundary rule: line breaks first; sentence and status cues only when a transcript arrives as one block. Original event labels remain unchanged.

Computed evidence

Run timeline

WAITING FOR INPUT

No run analyzed yet.Use the bundled sample or paste a transcript to reveal individual events.

Transcript evidence: order, reported duration and open signals

Your browser does not support video.

Read the explanation

The saved sample contains seven lines. Line breaks define seven ordered events before any type classification. Explicit Stage one and Stage two mentions give two distinct stage identifiers, even though those mentions do not establish stage completion. The two bars share sixty pixels per counted item. When input has multiple nonempty lines, the parser uses those lines. Only a single block invokes sentence segmentation and, when that yields one segment, cue-based splitting. It collapses internal whitespace in labels but retains the full original source text in JSON. Event IDs follow the original order. Classification checks duration first, then status, then tool actions, and finally progress. A line saying Updating image patch is a tool entry, while a Stage two unfinished line is status. Classification is evidence interpretation, not execution of the recorded actions. The saved duration line Working for six minutes four seconds contributes three hundred sixty-four seconds. The summary sums parsed minutes and seconds from every event, not only events classified as duration. Adding a separate report of two minutes would therefore yield four hundred eighty-four seconds. Both bars share one pixel per second. Repeated cumulative reports would double-count overlapping elapsed time, so this total cannot prove actual wall-clock runtime. The pattern uses the first matching minute value and first second value per line, rounds their sum, and does not interpret a clock-style six colon four as minutes and seconds. Tool evidence counts Compromise lexical terms, not billing tokens or genuine tool calls. Native timeline displays preserve labels and evidence, but no text parser can verify whether a logged command really ran. Stage two working tree is unfinished and Tests pending are the sample two blocker labels. Pending, unfinished, failed or blocked anywhere in a label marks the run open. This is keyword matching, so a sentence saying no tests pending would still trigger pending. An open signal asks for follow-up; it does not prove execution failure. Search and event-type filters change visible rows while summary and exports retain the complete canonical result. Grouping by event type changes display order within the grouped view; original IDs and JSON event order remain intact. JSON includes source text, terms, durations, stages and blockers. The HTML export includes all event labels and types with the summary but omits raw source and term details. The saved offline dependency is absent, so its initialization error and inert native controls are preserved separately. No successful parser execution, service call or backend request is claimed.