jazzyalex/agent-sessions v4.7

Agent Sessions 4.7

on GitHub

What's New in 4.7

Highlights:

• Kimi Code is the eleventh agent Agent Sessions reads. Its sessions land in the session list, in search and in Analytics, with tool calls and assistant output rendered in the transcript, a filter pill in the toolbar, and both Copy Resume Command and terminal resume. Agent Sessions switches it on when it finds Kimi Code on your Mac, and Settings → Agents gains a Kimi Code pane for the binary path and storage root.
• Active Claude cloud sessions appear in the Quota Meter. Sessions running on Anthropic's infrastructure were invisible outside Claude Desktop; they now sit in the pinned Quota Meter beside your local agents, marked Cloud and ordered above them. They carry no burn rate — there is no local transcript to measure — so they report status instead of a runway. Off by default, and offered only once you have saved a claude.ai cookie under Settings → Usage Tracking, since the fetch depends on it.

Features:

• Agent runs with no terminal now reach the Quota Meter. A claude -p run started by launchd, a shell script, or another app has no controlling terminal, and presence discovery required one — so those sessions never appeared in Live Sessions even though the main window indexed their transcripts normally. They are picked up now.
• Analytics counts Pi sessions. Pi has been a supported source since 3.8 but was missing from the analytics rollups. Its sessions, messages, durations and projects are now included, with a Pi-only filter alongside the others.

Improvements:

• Turning an agent on or off updates the rest of the app. Agent Sessions only watched five of the eleven agent switches, so toggling any of the other six left Analytics and the usage meters running on the previous set until relaunch.
• A refused Copy Resume Command leaves your clipboard alone. It cleared the pasteboard before working out whether it had anything to write, so the cases where it declines — a Codex session whose id cannot be resolved, a session with no known working directory — wiped whatever you were holding and replaced it with nothing.

Bug Fixes:

• Pi and Cursor sessions get their working directory back. After a full parse, a Pi or Cursor session lost the directory it ran in. Two things broke as a result: its resume command came back without the cd, so the resumed agent started wherever your terminal happened to point, and the search index stored no path for it at all, so filtering by project silently stopped matching those sessions. Both are fixed, and sessions indexed before the fix are re-read once in the background after you update, so the repair reaches the history you already have rather than only new sessions.
• A resume that fails now tells you. Every reason a resume could fail — a CLI that could not be found, a Warp that refused to launch, a session whose working directory is unknown — produced an error nothing ever displayed, so all of them looked identical on screen: nothing happened. Failures now come back as an alert. Warp in particular used to report success even when it never opened, and a failed launch would quietly retry a different command that could resume an entirely different session while reporting the one you clicked.
• Checking an agent's CLI no longer freezes the window. Probing a binary spawns a login shell and a --help for each candidate path, and eight of the agent coordinators ran that on the main thread with no timeout — so a slow shell profile or an unresponsive binary hung the whole app until it finished. Probes now run off the main thread.
• The Codex meter no longer sits on "reconnecting…" forever. With no Codex CLI run in the previous two days, the background refresh returned before ever making its live usage call — so whenever Agent Sessions was not the frontmost app the meter stayed on "reconnecting…" indefinitely, and only recovered if you focused the window. The live fetch no longer depends on a recent local session file. And when a fetch genuinely comes back empty, the meter now says no active session rather than implying a connection it is still waiting on.
• Codex sessions containing an auto-review turn reappear in the dollar view. Codex bills real tokens for its codex-auto-review model, but no price key matched it — and an unpriceable slice makes the whole session unpriceable, so the runway dropped those sessions from the $ view entirely and fell back to tokens per hour without saying why. The model is priced now.
• Antigravity code edits and web searches read as tool work again. CODE_ACTION and SEARCH_WEB carry the same result envelope as the file and command events, but fell through to the parser's default branch and rendered as system noise — a code edit was indistinguishable from a checkpoint. Transcripts the CLI truncated upstream are also marked as such now, instead of rendering as though they were complete.

Full release notes: https://github.com/jazzyalex/agent-sessions/releases/tag/v4.7