The session timeline
One Claude Code session ran the ET-SoC-1 work for eleven days. It started with the owner’s first message at 09:19 on Saturday 19 September and this snapshot ends at . Four things happened at once, and the chart below puts them on one time axis. The owner sent messages. The agents were the main agent and subagents, most of them started by workflow runs. The cards were four ET-SoC-1 cards in three lab machines. The artifacts were deploys of published pages and commits to the repository.
Each of the owner’s messages has a summary of at most 12 words, and its own words: hover over its mark, or tap it, for their start; click it, tap it again or press Enter to read it whole. Where a message told how the lab’s machines are reached, gave an address, named another person or linked a page that is not public, that part is replaced by a marked note. One report for the lab lead appears without its name, content or link.
1. Two weeks on one time axis
On a phone: swipe sideways to move along the time axis (it glides on when you let go), pinch to zoom, and swipe up or down to scroll the page. With a mouse or trackpad: drag or swipe sideways to move, Ctrl or ⌘ with the wheel or a trackpad pinch to zoom where the pointer is, and Shift with the wheel to move. The buttons zoom and move too. Tap, hover or tab to a mark for its details; on a message of the owner’s, click, tap again or press Enter to read its words; the numbered flags are the highlights of §2: tap a flag to read it and tap it again, or click it, to zoom there. On the time axis, ← → move, + − zoom and 0 shows it all.
2. Highlights
The turning points across all four tracks, in order. Click one to zoom the timeline there.
3. What the owner asked
Every message of the main session, in order: what was asked, as a summary of at most 12 words, and under it the start of the message’s own words; “read it whole” opens them all, and ← → step through every message. The time, the kind of message and its request number in the knowledge base are small labels; the estimate of the time spent reading and writing it is smaller still. Click a summary to see it on the timeline. The owner’s messages to the other sessions follow the list, a section for each, apart from its totals.
4. The totals
Who did the time
Every total below is taken from the timeline. The totals are in different units, so read them side by side and do not add them: the owner’s hours are estimated from what was read and typed, agent time is agent-hours (two agents busy for an hour count 2), and card time is card-hours. The session’s wall clock from the first message to the snapshot is .
Card-hours per card
How long the owner’s messages took estimate
Tokens
Day by day
Each row is one day, from midnight to midnight. From the top, the thin lanes show the owner’s messages, the main agent’s busy time, how many subagents were busy (scaled to , as the peak of the agents that did not fail is ; failed starts are drawn in grey on top and cut, as in the minute the usage limit hit on 20 September, when were counted), and how many cards were in use (up to 4). Click a row, or press Enter on it, to open that day in the timeline.
5. How it was counted
Data: docs/reports/data/2026-09-27-session-timeline, which holds the seven extracts and timeline.json,
which this page draws. Generator: tools/timeline/build_timeline_data.py. Extraction scripts:
tools/timeline, whose README.md gives their order and settings. They read the session’s transcripts and the lab machines’ logs, which are not in the
repository, so the committed extracts are the page’s inputs; tools/timeline/sanitize_extracts.py removes from
them what the page withholds and checks that nothing private is left.
Related reports
- Limits of observability: the hub, which lists every report this week produced.