gigamanage weekly downloads statistics

gigamanage

Browse, search and resume your AI coding agent sessions — across Claude Code, Codex, and whatever you use next.

Agent sessions pile up faster than your memory of them does. After a few weeks you have hundreds of transcripts, and the built-in pickers sort them by time and label them with a title generated in the session's *first* few seconds. That title tells you where the work started. When you're deciding what to pick back up, you need to know the latest status.

It's like saying "good morning" to your agents

gigamanage is a small CLI, gm, that helps you reorient yourself among your agent sessions.

gm shows a full context card for the highlighted session alongside it — what landed, what's still open, and the next concrete step. Hit enter and you're back in the session, in the right harness and the right directory. ctrl-r reloads the list to your most recent sessions and starts summaries for any that need one — handy when you left the picker open while an agent was working. (Without fzf, the numbered list takes r for the same thing.)

Claude SBS

Both of these are looking at one webshop repo with six recent sessions. The built-in picker labels each one with the title Claude Code generated in its opening seconds; gm ls labels it with where the work actually ended up.

claude --resume gm ls

"webhook retries are flaky" is what you asked for four hours ago. Whether it got fixed is anyone's guess — and the two sessions that died mid-task look exactly like the four that didn't. The retry fix landed but the timestamp check never got written, and the Node 22 bump left the build red. Both are flagged ⚠: they ended mid-task.

What makes it different

Summaries describe what the work became. gigamanage reads each transcript's *arc* — what you originally asked for, how the work moved, your last instructions, the agent's final message, the files it touched, the last command that failed — and writes four things: what the session is about, what landed most recently, what's still open, and the next concrete step. The harness title names the opening prompt and never revises it; this tells you where the work actually is. That's the whole point of the tool.

It knows when work was cut off. Sessions that ended mid-task are flagged ⚠. Those are usually the ones you're looking for.

It works across harnesses. Claude Code and Codex today, with one small interface for adding more. gm resume hands off to the right CLI — claude --resume or codex resume — in the session's original directory.

Agents can use it too. Every read command takes --json. Your agent can shell out to gm grep "flaky test" --json to find what you already tried, instead of asking you.

A live view of every agent at once. Drive your agents in tmux and gm keeps every pane's border labelled with what its session is doing (alt-g), and overlays the full cards in place on demand (ctrl+g) — so you can glance across all of them and drill in only where it's worth it. Live from tmux has the setup.

Install

From npm (recommended):

Or run it without installing anything:

From source — for hacking on it, or to run an unreleased commit:

With npm link, gm tracks your working copy: re-run npm run build and the next gm picks it up. To run straight from TypeScript without building, use npm run dev -- ls. To unlink later: npm unlink -g gigamanage.

Requires Node 20+. Three optional companions, all surfaced by gm doctor:

• ripgrep (brew install ripgrep) — needed for gm grep. • fzf (brew install fzf) — upgrades the picker to fuzzy search with a preview pane. Without it you get a numbered list. • tmux 3.2+ (brew install tmux) — unlocks the live alt-g label agent and the ctrl+g peek overlay (see Live from tmux below).

Summaries are written by a model, so the first time you run gm it asks which one to call — Claude Code, Codex, anything that reads a prompt on stdin, or nothing at all. Change your mind any time with gm setup. GIGAMANAGE_SUMMARY_CMD='codex exec' overrides it for a one-off, and nothing prompts when the output isn't a terminal, so gm ls --json stays safe to script.

Usage

Summaries are cached and only regenerate when a session actually changes, so you pay for each one once.

By default the list hides two kinds of noise: subagent transcripts (--include-sidechains) and non-interactive runs like claude -p or codex exec (--include-automated).

Ask across your sessions

A list answers "what was I doing?" one row at a time. gm ask answers the question that spans them: *given all of it, where should I be looking?*

It starts from the summaries already on disk — so it costs one model call, not a scan of half a gigabyte. When the summaries don't carry enough, it runs gm grep against the real transcripts and reads what you actually said.

In the picker, ctrl-o opens it on the session you're highlighting: ask "what's left here?", read the answer, and land back in the list exactly where you were. (Without fzf, the numbered list spells it a.)

It isn't shift+f because fzf's query line eats plain letters — typing F types an F. And it isn't alt-a because macOS sends å.

Summaries write themselves

Every gm command keeps the sessions you just looked at summarized. gm ls shows 20 by default, so it keeps 20 written; gm ls -n 50 keeps all fifty. Any that are missing or stale are handed to a detached background process, eight at a time, and the command returns immediately:

| marker | meaning |

| ◐ | being summarized right now | | ○ | no summary yet, and nothing running | | ⚠ | the session ended mid-task — usually the one you want |

The foreground command never waits on a model: it prints and exits, and the summaries appear on your next run. Only one background pass runs at a time — a lock in ~/.cache/gigamanage means five gm ls in a row start one summarizer, not five. A pass writes at most 50; the rest are picked up next run, and it says so.

The notice goes to stderr, so gm ls --json stays clean for agents and pipes.…

Description from the npm listing.