Skip to content

feat(session): idle breaks, room for the lists, readable timeline and shell - #40

Merged
byronwall merged 6 commits into
mainfrom
session-overlay-10x
Nov 22, 2026
Merged

byronwall merged 6 commits into
mainfrom
session-overlay-10x

Conversation

@byronwall

@byronwall byronwall commented Nov 22, 2026 •

Copy link
Copy Markdown
Owner

The session overlay now gives the lists more room and is easier to read. Idle gaps show as breaks in the chart and the timeline, with clock times on both sides. Timeline rows show the kind of each tool and what it touched. The agent picker describes each subagent. Shell commands show one statement per line, with colors.

What you'll see

  • Idle gaps. When no request runs for longer than the 5-minute prompt cache, the chart bars break with a hatched gap that names its length (9m, 4h12m). The gap is amber when the cache expired and gray when it held. Clock labels sit on both sides of each break. The timeline adds a striped row in the same place, such as "9m 08s idle · the prompt cache expired, so the next request wrote 78.7K tokens again".
  • More room. The chart runs to the top of the overlay. The title, meta line and totals (in four columns) sit on its left, so the separate header row is gone. The bar above the chart starts with Review with an agent and Open in Claude, then the chart buttons (widen to full width at 200px tall, put back, hide), then the key, with × at the far right. The browser keeps the chart choice. On a real 63-request session at 1300px, the list starts at 229px from the top, or about 120px with the chart hidden. Before this PR it started at about 365px. The toolbar never wraps.
  • Highlighting. A selected finding shows as "Highlighting ×" in the chart bar, tinted like the rows it lights. Before, a lightning-bolt chip in the toolbar read like a filter.
  • Your messages only. A toolbar button on the Timeline tab keeps just your messages, each with the agent's last word for that turn, so the list reads as the conversation. The inspector opens on the turn.
  • Timeline structure. Requests are one line where they fit. Said holds what the agent said (two lines at most, Markdown marks removed) and Did holds the tools. Each turn is a one-line header with your message on the left and its facts on the right. Times drop AM/PM in request rows. Each tool has an icon for its kind (read, edit, shell, search, subagent, web, MCP). A monospace hint shows what the tool touched: the programs a shell call ran (git head tail, never the text of a heredoc), file names, or a search pattern. Numbers are dim unless they matter: a cache share under 50% is amber and bold, and new input over 20k is bold. In All agents, each row and bar shows its agent's color.
  • Agent menu. The native select is now a button that names the agent in scope ("Main agent +2 subagents"). Its menu shows each subagent's description, kind, model, requests, calls, failures, duration, cost, and a bar of when it ran. Up and Down move through it, Enter picks, and Esc closes only the menu.
  • Previews, not text dumps. The inspector's call lists show each call as time, then content, then duration. The content is the tool, plus "in " for a leading cd, above a two-line preview. More opens the rest in place, and short commands take one row. Finding details show two lines in the list and four in the inspector. Calls are named by their command (without the leading cd), file, pattern or title, never raw JSON. "Took over a minute" lists the three slowest first.
  • Long sessions. I tested a 14-day automation session with 1,221 requests, 349 messages and 355 idle gaps.
    • Breaks share at most a fifth of the chart: the longest gaps stay as breaks, and the rest become ticks on the axis.
    • Clock labels add the date wherever it changes, and dense messages draw as thin ticks.
    • Runs of one polling check now make one finding, even when its arguments vary ("Ran the same ssh call 723 times in 23 runs · 22 variants"). That session's findings went from 31 to 10.
    • Codex code-mode calls (tools.exec_command({cmd:…})) now show their commands instead of JavaScript.
    • Automation turns read as "Automation · ", and each reply as Quiet or Notified with its message. A prompt that repeats an earlier turn's takes one line, marked "same as turn N".
    • Selecting a row no longer rebuilds the 1,900-row list and the 1,200-bar chart. Each selection took about 490ms before and takes about 15ms now.
  • Shell commands. In the inspector and in the per-tool call lists, a one-liner is split at its top-level ;, &&, || and |, and loop and if bodies are indented. The split still runs the same command. Commands, flags, strings, paths and operators get colors, and a cd line is dimmed.

Proof

Before After
before timeline after timeline

The same session at 1440px. In the after shot, the list starts about 75px higher. The 9m idle gap (10:13 → 10:22) is a labeled amber break, and the timeline rows show tool kinds and file names.

agent menu
The agent menu with both subagents described and their run bars on the session clock.

main agent with idle rows and Bash
The main agent alone. It has two breaks: 8m, where the cache held (gray), and 9m, where it expired (amber). The matching striped rows are in the timeline, and a Bash call in the inspector is split and colored.

Bash call list
The Tools tab with Bash selected. Each call is one row when it is short and a two-line preview when it is long, and its command is formatted and colored.

1000px wide
At 1000px, the agent column becomes a color dot, so the number columns stay in view.

Checks:

npm test   # all files pass, including the new shell.test.js

shell.test.js protects the command split. It checks that the split never breaks inside quotes, $( ), a subshell, or a heredoc, and that it indents loop bodies.

Limits

  • The long-session changes have no screenshot. The only long session I could test is a private log with a real server address. The numbers above come from it, measured in the browser.

  • Gaps use a fixed 5-minute threshold. Sessions with 1-hour cache writes show many gray "cache held" breaks. They are real jumps in time, but they cost nothing.

  • The chart key is hidden on phone widths.

  • Your messages only is not kept in the URL. A refresh shows the full timeline again.

🤖 Generated with Claude Code

byronwall and others added 6 commits November 21, 2026 23:15
…nd shell

- Chart breaks at idle gaps longer than the prompt cache, names the gap, and
  labels the clock on both sides; amber when the cache expired.
- Totals fit on one line, the chart is shorter and can be hidden (kept per
  browser), and the key moves to the toolbar.
- Timeline rows get texture: your messages stand out, tools show their kind
  and what they touched, numbers dim unless notable, idle rows sit between.
- The agent picker is a menu that describes each subagent and when it ran.
- Shell commands split into statements and get colors (public/shell.js).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…dings

- Inspector call lists lay each call out as time | content | duration, with the
  tool and an "in <folder>" label for a leading cd above a two-line preview
  and a More button that opens the rest in place.
- Finding details clamp to two lines in the list and four in the inspector.
- Call summaries use the command without its leading cd, files, pattern or
  title, never raw arguments; slow calls list the three slowest first.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…filter

- Totals sit in two columns beside a shorter chart (112px). The chart's own bar
  holds its key and buttons to widen it (full width, 200px) or hide it.
- The toolbar no longer wraps, so the findings chip can't add a line.
- "Your messages only" keeps just your messages, each with the agent's last
  word, and the inspector opens the turn.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… the chart bar

- The chart runs to the top of the overlay; the title, meta and totals (four
  columns) sit on its left. The bar above the chart starts with Review and
  Open, then the chart buttons and key, with close at the right.
- A selected finding shows as "Highlighting <finding>" in the chart bar, in
  place of the toolbar chip that read like a filter.
- Timeline rows split into Said and Did columns, one line where they fit;
  turn rows are one line with the facts on the right; times drop AM/PM.
- Shell hints and colors skip heredoc bodies (test added).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… line up

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Chart breaks share at most a fifth of the plot: the longest gaps stay as
  breaks, the rest become ticks. Multi-day sessions get dates on the clock
  labels; dense messages draw as ticks; the window label no longer collides.
- Repeat findings merge runs of one check across the session, even when its
  arguments vary (31 findings became 10 on a heartbeat session).
- Codex code-mode calls (`tools.exec_command({cmd})`) read as their commands,
  also when the log cuts the arguments short.
- Automation turns read as "Automation · <name>" and replies as Quiet or
  Notified with their message; repeated prompts collapse to one line; idle rows
  the cache rode out show only from 30 minutes.
- Picking a row moves the highlight instead of rebuilding a 1,900-row list and
  1,200-bar chart: ~490ms became ~15ms.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@byronwall
byronwall merged commit bd16320 into main Nov 22, 2026
1 check passed
@byronwall
byronwall deleted the session-overlay-10x branch November 22, 2026 04:54
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant