jazzyalex/agent-sessions v4.8

Agent Sessions 4.8

on GitHub

What's New in 4.8

Highlights:

• Grok CLI is the eleventh current agent source Agent Sessions reads. Its sessions land in the session list, in search, in Analytics and in the Image Browser, with tool calls, reasoning 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 Grok CLI on your Mac, and Settings → Agents gains a Grok CLI pane for the binary path and storage root. Subagent runs are shown under the session that spawned them — Grok records that parentage in the parent's own tree, so grok sessions list does not show them at all. Schema verified against 141 real sessions. Contributed by @thedavidweng in #55.

Features:

• Analytics counts every agent you have enabled. The rollup ran on a hand-written list of sources, and Cursor and OpenClaw were never on it — their sessions were indexed, searchable and visible in the list, but contributed nothing to any total, chart or project breakdown, and nothing reported the gap. Analytics now derives from the full set of sources, so every agent counts and a new one cannot be forgotten. Cursor also gains the agent-picker entry it never had. Existing installs rebuild their analytics once on the next launch to pick up the two missing sources; it reads the session index and touches no files.

Bug Fixes:

• An unrelated binary earlier in your PATH no longer masks the real CLI. Cursor, Kimi and Pi searched candidate locations group by group and returned the first executable they found, even when it did not advertise the flags the CLI needs — so a same-named binary in the login shell's PATH won before the actual install directory was ever examined, reporting its version as the agent's and refusing resume outright. Each now searches one ordered list, and a binary that advertises the right flags always wins.
• Analytics no longer reports "no sources enabled" over data it has. Its empty state, the banner warning that totals exclude an agent, and the agent picker were three more hand-written lists, each stopping at a different source. With Cursor or OpenClaw as your only enabled agent, Analytics claimed nothing was enabled while showing that agent's data, and the picker could not isolate either. All three lists now derive from one source of truth.
• The Image Browser credits the prompt that produced an image, not the scaffolding around it. Agents inject housekeeping blocks into the conversation as if they were your messages — an AGENTS.md preamble, <environment_context>, a <system-reminder>. The transcript already knew to look past those when deciding which prompt an image belongs to; the Image Browser did not, so an image could be filed under a <system-reminder> there while the transcript filed the same image under the real prompt. Claude injects those blocks mid-conversation, so this was not limited to the opening of a session. Both now agree.
• "Live sessions only" includes Claude Desktop chats. Claude Desktop sessions have no terminal process for the usual live probes to find, so the filter could hide running Claude chats even though the Quota Meter saw their runway rows. The live session engine now synthesizes Claude Desktop presences from the same runway identities, preserving active/idle state and avoiding duplicates when a transcript is already claimed by another probe. The approach — Claude Desktop chats as first-class live presences — came from @Krazycatt in #51; my thanks.
• The Session Runway switches units immediately. Selecting Wk could leave the previous tokens-per-hour values visible while the asynchronous runway scan recalculated, and an automatic fallback after a quota window appeared or disappeared had the same problem. Rows now change to the effective unit at once and show a waiting value until recalculation finishes.
• Codex plan upgrades recover without a relaunch or unnecessary sign-in. If you hit the weekly limit and then upgrade your plan, the running app can still hold its old failed-fetch cooldown and claim the saved login needs attention even though it remains valid. Agent Sessions now clears that process-local state and retries silently before showing Fix…; it only recommends codex login when the clean retry also rejects the saved credentials.

Full release notes: https://github.com/jazzyalex/agent-sessions/releases/tag/v4.8