Claudoscope v1.4.0

The session viewer release. Until now the Sessions rail showed a transcript and left you to scroll. 1.4.0 turns it into a place to study a session: a Trace tab lays the whole run out on a time axis with a Replay that plays it back as a movie, a minimap above the chat shows the session's shape at a glance, bookmarks pin the turns worth coming back to, and a Resume button drops you back into the session in your terminal. Also cuts main-thread CPU while sessions stream (43 percent down to under 1 percent with no window open), adds Haiku 5.5 pricing, and makes the menu bar live-session badge draw as designed.

Requires a one-time full reparse on first launch (parser version 14).

New Features

• Trace tab. A session viewer tab that lays the whole session out on a time axis: your prompts, how long Claude worked on each, every tool call by category with errors and blocked calls marked, subagents forking off and merging back, context fill with a cut at every compaction, and cumulative cost. Idle stretches over five minutes collapse to a labeled break so a two-day session fits one screen (toggle off for true time). Hover for a readout, click any mark to open it in Chat, drag to measure a range (turns, tool calls, failures, spend), arrow keys step turns.
• Replay. A play button in the Trace tab sweeps a playhead across the session at 60×, 300×, 1200× or fitted to twenty seconds, revealing the lanes as it goes with the current turn, context fill and spend so far in the footer. Space toggles, Escape stops.
• Share card. Export the trace as a 1200×630 PNG (save or copy) with the project name, duration, turns, tool calls, agents, compactions, cost and model family. No transcript text and no agent descriptions go into the image.
• Session minimap. A strip above the chat draws the whole session's anatomy: user and assistant turns, tool calls colored by category, tool errors, blocked actions, compactions, subagent spawns and bookmarks. Click a mark to scroll the chat to that record.
• Bookmarks. Right-click a user or assistant turn to bookmark it, optionally with a note. Bookmarks persist in the local SQLite cache and show as a sidebar badge on the session, an accent on the turn in Chat, a mark on the minimap, and a cross-session Bookmarks section in the sidebar. The command palette (Cmd+K) opens them too.
• Resume in terminal. A Resume button in the session header opens a new window in your running terminal (Ghostty, iTerm2, or Terminal.app as the fallback), in the session's working directory, and runs claude --resume <id>. Worktree sessions resume in the worktree, not the main checkout. The session id is validated before it reaches a shell line, and the button is disabled for subagent transcripts.
• Menu bar live-session badge (optional). Settings, Appearance gains "Show live session count in the menu bar", off by default. When on, a rounded badge next to the icon counts running sessions: blue normally, amber when any of them is waiting on you, red once a wait escalates. The 1.3.0 waiting count showed as a bare digit because MenuBarExtra never drew its capsule; the label is now rendered as one image so it draws as designed. The monochrome setting shows a plain bold digit instead.
• Haiku 5.5 pricing. claude-haiku-5-5 bills at $0.10/$0.50 per million tokens with $0.01 cache reads. It is the first current model with a prompt-length tier: a request over 100K tokens (input plus cached prefix) bills at $0.50/$2.50 for that message, and the estimator now decides this per message from the actual request size. Before this it was priced as Haiku 4.5, ten times too high on the common tier.

Changed

• No more red dot for bypass sessions. The menu bar icon no longer turns its dot red while a session runs with skipped permissions; the Fleet card still carries the Bypass tag. The red dot now means an unseen cost alert only.
• Stable download URL. The download counter at dl.claudoscope.com gains /latest/Claudoscope.dmg, which resolves the newest release from GitHub (cached five minutes), counts the download, and redirects like the versioned route. Used by the website's download buttons.

Fixed

• Main-thread CPU while a session streams. While any Claude Code session wrote its transcript, Claudoscope sat at 30 to 50 percent of a core with memory sawing by up to 170 MB per burst, because every append re-ran work that was linear in the whole corpus: SwiftUI deep-compared every session summary in the sidebar, the popover walked the corpus six times per render, ISO timestamps went through ICU, and the chat re-derived its turn data per assistant row and again per append. Sidebar rows now take a small row model, timestamp parsing has an integer fast path, the popover's lists are rebuilt once per append, fleet and analytics rebuilds are coalesced to once per second, and chat data is derived once per session. Measured with a session streaming: under 1 percent with no window open and about 9 percent with the dashboard on a live chat, down from 43 percent in 1.3.0.
• Popover left the process busy after closing. MenuBarExtra keeps the popover's view hierarchy alive after it closes, and the pulsing dot on the Active Sessions card kept rendering off-screen at 11 to 14 percent CPU. Animations now pause while the popover window is hidden; idle is 0 percent again.
• Dashboard memory. The sidebar realized every row of every expanded project up front, so a window showing twenty-five rows held about 2,200 text lines and 2,300 rasterized symbols. Rows are now realized as they scroll into view. The same window dropped from 428 MB to 195 MB.
• Exited sessions stuck at "Finished, ready for review". A Stop hook left a session in the attention queue (and the menu bar count) until its transcript moved or it reported busy again. Quitting claude did neither, so the session sat there for the whole 24-hour window. The pending hook event is now dropped when the process leaves Claude Code's registry; a Stop hook newer than the last busy report still marks the exit as clean rather than a mid-turn crash.
• Sonnet 5.5 cache reads halved. claude-sonnet-5-5 matched the Sonnet 5 rate row by substring and billed $0.20 per million cached tokens; its own rate is $0.10. Input, output and cache-write rates were already right.
• Mythos 5.1 cache reads. claude-mythos-5-1 billed Mythos 5's $1.00 cache reads; its own rate is $0.25, the same split Fable 5.1 made against Fable 5.
• Rate markers are ordered most-specific first. A new rate split for a point release (sonnet-5-5) can no longer be captured by its parent's marker (sonnet-5).