@gigaflow/gmux 0.15.0

package/CHANGELOG.md

@@ -0,0 +1,654 @@
+# Changelog
+
+Notable changes, newest first. Versions follow [semver](https://semver.org): while
+0.x, a **minor** bump means behavior changed in a way you should read about before
+upgrading, and a **patch** is a fix that asks nothing of you.
+
+## 0.15.0
+
+### Always-on workspace awareness — daemon, cockpit, memory guardian
+
+gmux gains the layer the name always implied: a resident daemon that watches your
+whole tmux workspace and tells you, at a glance, what every pane is doing — without
+focusing each one in turn. It builds on the existing sensing (transcript adapters,
+summaries, border labels) and is **additive** — nothing about the picker changed.
+
+- **Start it.** `gmux daemon` runs a resident daemon that senses every pane (agent
+  and plain shell) on a ~1.5s tick and serves one workspace model over a local
+  socket (with a snapshot-file fallback). `gmux tmux install` wires the border
+  labels and the `ctrl+g` cockpit into your `~/.tmux.conf`.
+- **Ambient border labels.** Each pane's border shows a state glyph — ● working ·
+  ◔ waiting · ✗ error · ✓ done · ○ idle — plus a one-liner: instant heuristic state
+  always, and an LLM-written "what it's doing" line a beat later (change-gated and
+  debounced, so it stays cheap).
+- **The cockpit (`ctrl+g`).** A full-workspace grid: per-pane state, memory, a
+  one-liner, and last activity, with a ranked memory column that names the hog
+  (and surfaces memory it can't attribute honestly) and a guardian log up top.
+- **Memory guardian.** Under host memory pressure the guardian can broadcast a
+  "checkpoint your work and pause non-essential tasks" message straight into your
+  agent panes, naming the top consumer. This is the one action gmux takes on your
+  behalf, so **`gmux setup` discloses it and asks for consent** — default `auto`,
+  with `notify` and `off` offered. It only ever types into known agent panes (never
+  a shell where a human is at the keyboard) and honors a cooldown so it can't spam.
+
+Two lanes, one tool: the cockpit answers *"what's happening right now across my live
+panes"*; the picker (`gmux`, `gmux ls`, `ctrl+shift+g`) still answers *"browse and
+resume any session across time."*
+
+The fast path — pane state, memory, and the guardian — is built to survive anything
+the LLM or tmux does: a hung summarizer only lets labels go stale, and a tmux hiccup
+idles a tick; neither freezes triage. See [`docs/gmux.md`](docs/gmux.md) for the
+design, and the memory-attribution caveats (RSS ranks reliably but isn't an exact
+total; detached/containerized children show as *unattributed*).
+
+## 0.14.0
+
+### Renamed to gmux
+
+The project is now **gmux** — "giga multiplexer": tmux, but LLM-native. This is a
+clean break from the old `gigamanage` / `gm` names, so it asks a few things of you
+on upgrade:
+
+- **Install the new package.** It publishes as `@gigaflow/gmux` and installs a
+  single command, `gmux`. The old `gm` and `gigamanage` commands are gone.
+  ```bash
+  npm uninstall -g gigamanage        # remove the old one
+  npm install -g @gigaflow/gmux    # gives you `gmux`
+  ```
+- **Environment variables are now `GMUX_*`.** `GIGAMANAGE_SUMMARY_CMD` →
+  `GMUX_SUMMARY_CMD`, `GIGAMANAGE_AUTO_SUMMARIZE` → `GMUX_AUTO_SUMMARIZE`, and so
+  on. The old names are no longer read — update your shell profile.
+- **State moved.** Config now lives in `~/.config/gmux` and the cache in
+  `~/.cache/gmux` (previously `…/gigamanage`). Nothing is migrated: run `gmux setup`
+  to choose your provider again, and summaries regenerate on first use. You can
+  delete the old `~/.config/gigamanage` and `~/.cache/gigamanage` directories.
+
+Everything the tool *does* is unchanged — same picker, same summaries, same
+cross-harness resume, same tmux overlay. Only the name and its surfaces moved. The
+[design doc](docs/superpowers/specs/2026-08-11-gmux-design.md) lays out where gmux
+goes from here.
+
+## 0.13.6
+
+### Summaries follow the pane, even after you move panes around
+
+When several panes share a directory and their agents were started fresh
+(`claude --model=…`, no `--resume <id>`), gigamanage had nothing to tell the panes
+apart — no session id on the command line, no `gm run` link — so it handed out the
+newest sessions in the directory in pane-list order. Move a pane, or let a session
+update, and the summaries reshuffled onto the wrong cards.
+
+The overlay now pairs those panes to sessions by **process start order**: within a
+directory, the oldest agent process takes the oldest-started session, and so on.
+A process's start time is stable and belongs to the pane's real process, so the
+mapping holds when you move, swap, or join panes. A directory with a single agent
+is unchanged. Exact matches — a `--resume <id>` on the command line, or a `gm run`
+link — still win outright.
+
+## 0.13.5
+
+### ctrl-g closes the overlay it opened
+
+The `ctrl-g` peek now toggles: press it again to dismiss, the same as the key that
+opened it. Before, only Esc (or ctrl-c / ctrl-d) closed it — once the ask box
+landed, every other key typed into that box, so a second ctrl-g was swallowed and
+the overlay stayed up. Esc still closes; the ask-box hint now reads `^G/Esc close`.
+
+## 0.13.4
+
+### Codex summaries work again after an upgrade
+
+A config written by an older `gm` froze the codex command as `codex exec`. When
+the catalog later gained `--skip-git-repo-check` (so codex can summarize outside a
+trusted git checkout), that flag never reached an existing config: summary
+resolution replayed the stored command verbatim, so every codex summary failed
+with *"Not inside a trusted directory and --skip-git-repo-check was not
+specified"* and the `ctrl-g` overlay stayed stuck on "no summary yet".
+
+A **known** provider now re-derives its argv from the catalog rather than the
+command frozen at setup time — the same self-healing `gm ask` already had. A
+`custom` provider is still run exactly as you wrote it. No re-run of `gm setup`
+needed; upgrade and codex summaries resume.
+
+## 0.13.3
+
+### A fresh claude pane no longer shows a codex session (or vice versa)
+
+Resolving a fresh session — one with no id on its command line — now reads the
+harness from the agent's own argv (`claude` vs `codex`) and hard-filters to it, so
+it can only match a session of the SAME harness in its directory. Before, a fresh
+`claude` pane could grab the newest codex session in the repo (the harness guess
+came from `pane_current_command`, which is just `node`). If no session of that
+harness is known for the directory, the pane shows nothing rather than the wrong
+one.
+
+## 0.13.2
+
+### Force a refresh, and no more cross-window duplicate panes
+
+- **Force refresh:** press **ctrl-r** in the `ctrl-g` overlay to regenerate the
+  visible panes' summaries now, ignoring the divergence gate. From the shell:
+  `gm summarize <id> --force`, `gm summarize --recent 20 --force`, or
+  `gm summarize --all --force`.
+- **Cross-window fix:** the overlay (and `gm ask --window`) used to resolve only
+  the current window's panes, so a fresh agent with no session id on its command
+  line could grab the session another window's pane already owns — making two
+  panes show the same summary. They now resolve every pane in the server and keep
+  this window's, so the de-duplication holds across windows.
+
+## 0.13.1
+
+### The broadcast answer appends to the card, it does not replace the summary
+
+When you ask in the overlay, each card now keeps its summary and appends the
+per-pane answer (`▸ answer`, or `▸ asking…` while it lands) beneath it — reserving
+room so both show, instead of the answer swallowing the summary.
+
+## 0.13.0
+
+### The ask box broadcasts to every pane
+
+A question typed in the `ctrl-g` ask box is now answered **per session**: it fans
+out to each pane in parallel and every card shows ITS own answer, not one
+synthesized blob. "What is the single most urgent next action?" → each agent's
+card shows its own. While answers land, cards read `asking…`. Press Enter on an
+empty box to clear the answers back to the summaries; Esc (or ctrl-c) closes.
+
+## 0.12.0
+
+### The ctrl-g overlay has an ask box built in
+
+The `ctrl-g` overlay now carries a text box across the bottom. Just start typing a
+question and press Enter — the answer renders above the cards, fanned out over the
+current window's agents ("what is each doing?", "what is most urgent?"). Keep
+asking; Esc (or ctrl-c) closes. This replaces 0.11.0's "press a to launch a
+separate chat" with an integrated box on the overlay itself.
+
+## 0.11.0
+
+### Ask across the agents you are looking at
+
+In the `ctrl-g` overlay, press **`a`** to open a chat that fans out over the agent
+sessions in the current window. Ask high-level orienting questions — "what is each
+one doing?", "what is most urgent?", "which are waiting on me?" — and keep asking;
+it is `gm ask`, scoped to the panes in front of you rather than your whole recent
+list. Any other key still just closes the overlay.
+
+Also available directly: `gm ask --window <window-id>`.
+
+## 0.10.2
+
+### The ctrl-g cards say when they last updated
+
+Each card in the `ctrl-g` overlay now shows when its summary last landed —
+`updated 3m ago` (seconds, minutes, hours, days) — and while a refresh is in
+flight it reads `refreshing… · updated 3m ago`, so you still know how old the
+summary you are reading is.
+
+## 0.10.1
+
+### The overlay opens about twice as fast
+
+Resolving which session each pane runs walked the pane's process tree with a
+`pgrep` per node — dozens of process spawns down a deep agent tree (an agent with
+MCP-server children), which dominated the `ctrl-g` latency. It now takes a single
+`ps` snapshot and walks the tree in memory. Measured: the overlay path roughly
+halved.
+
+### A fresh agent no longer copies another pane's summary
+
+A pane running a *fresh* session (no session id on its command line) fell back to
+"newest session in this directory" — which is whatever another pane is actively
+working on, so its summary got copied onto the fresh one, especially across
+windows. Panes are now resolved together: an exact match (a `gm run` link or a
+session id read from the agent's argv) claims its session, and no heuristic pane
+may pick a session another pane already owns.
+
+## 0.10.0
+
+**Upgrading:** the summary prompt changed, so every summary regenerates on its
+next background pass (once, in the background). No action needed.
+
+### Summaries in three widening tiers
+
+A summary now narrows your attention gradually, so you decide how much to give a
+session:
+
+- **headline** — the scannable one-liner (what the `alt-g` label shows);
+- **overview** — 2-3 sentences framing it;
+- **summary** — new: a paragraph or two that actually reorients you, tracing how
+  the work evolved and where it stands.
+
+`ctrl-g` (and `gm show`) render all three — headline, overview, then the
+drilldown — above the recent/open/next status. The label stays the one-liner;
+the drilldown is one keypress away.
+
+### Fixes
+
+- `alt-g` off now fully clears the labels. Older versions set
+  `pane-border-status` per window, and that override kept the border (and its
+  headline) up even after the global was turned off. Toggling off now clears the
+  per-window override and wipes each pane's label.
+
+## 0.9.0
+
+gigamanage becomes what it was always meant to be: **a background agent that
+keeps you up to date on your agents' latest work.**
+
+**Upgrading:** re-run `gm tmux install` (or, on Oh My Tmux, keep the `Alt-g` line
+in `~/.tmux.conf.local`) and reload. `Alt-g` now toggles a live service rather
+than painting labels once. Nothing else changes.
+
+### `Alt-g` runs a live label agent
+
+`Alt-g` starts a single, lightweight background service (`gm watch`) that every
+few seconds resolves every agent pane across all your windows and keeps its
+border label current — the session's headline, where the pane is, with the
+content still visible. A pane whose summary is being regenerated shows
+`gm summaries loading…` until it lands. `Alt-g` again stops it. `ctrl-g` is still
+the full card when you want the detail; the two share one continuously-maintained
+cache, so the popup opens already current.
+
+You can also drive it directly: `gm watch` starts it, `gm watch --stop` stops it.
+
+### Summaries refresh when work *diverges*, not on every keystroke
+
+A live session used to be re-summarised on essentially every message. Now each
+summary carries a compact **SimHash fingerprint** of its content (16 hex
+characters, fixed size however long the session grows), and a session is
+re-summarised only when that fingerprint has drifted past a threshold — with a
+safety net: a new tool failure, a flip to ended-mid-task, or a change in the
+files touched always refreshes, so a small-but-important change is never slept
+through. Tune the threshold with `GIGAMANAGE_REFRESH_DISTANCE`; the loop interval
+with `GIGAMANAGE_WATCH_INTERVAL_MS`.
+
+## 0.8.2
+
+### Pane-border labels show on every pane, in full
+
+Two fixes to the `Alt-g` label HUD:
+
+- **Every pane, not just the active one.** The label inherited
+  `pane-border-style`, which themes like Oh My Tmux dim to near-background for
+  inactive panes — so every label but the active pane's was invisible. The label
+  now forces its own foreground colour and reads on all panes.
+- **No truncation.** The headline was capped at 60 characters; it's now stored in
+  full and tmux clips it to the pane's width at render, so a wide pane shows the
+  whole thing.
+
+A tmux pane border is a single line, so a headline wider than the pane is clipped
+rather than wrapped — press `ctrl-g` for the full card when you need it.
+
+## 0.8.1
+
+### Resolving a pane skips a needless `lsof`
+
+`gm tmux label` and the `ctrl-g` overlay read each pane's agent process to find
+its session. When the agent's command line already carries the session id (a
+resumed session — the common case), the id is exact and the process's working
+directory is never needed — but 0.8.0 looked it up anyway (`lsof` on macOS,
+~100ms per pane). It's now skipped unless there's no id on the line, so resolving
+a window of resumed agents no longer pays for a directory lookup per pane.
+
+## 0.8.0
+
+**Upgrading:** re-run `gm tmux install` to pick up the new `Alt-g` binding, and
+reload with `tmux source-file ~/.tmux.conf`. Everything is additive; nothing you
+rely on changes.
+
+### It knows which session each pane is running
+
+The overlay used to map a pane to a session by matching the pane's working
+directory — but that's the *shell's* directory, usually `~`, not the agent's, so
+panes resolved to the wrong session or to nothing. gigamanage now reads the pane's
+own process: the agent's command line carries the session id verbatim (`codex
+resume <id>`, `claude --resume <id>`), and that id *is* the session. Where there's
+no id on the line (a fresh session), it uses the agent process's real working
+directory instead of the shell's. No `gm run`, no setup — it just reads what's
+already there.
+
+### A label on every pane's border
+
+`Alt-g` toggles a one-line summary onto each pane's border — the session's
+headline, right where the pane is — while the pane content stays fully visible. It
+answers "what is each of these agents doing?" without taking over the screen; the
+`ctrl-g` full-card popup is still there when you want the detail. The label is
+stored in a pane-local option gigamanage owns, so a running agent's own title
+updates can't clobber it.
+
+### The peek does less work
+
+`ctrl-g` paints from the cached index and resolves each pane once, then upgrades
+to a full index read in the background. An open overlay no longer re-scans
+thousands of session files every second — it re-reads only the summaries that
+change.
+
+## 0.7.1
+
+Fixes and polish for the tmux peek overlay shipped in 0.7.0 — the first release
+made it usable in practice.
+
+### The `ctrl-g` binding now actually opens the overlay
+
+`gm tmux install` wrote `gm overlay #{window_id}`, but tmux does not expand that
+format inside `display-popup -E` — so the shell saw the `#` and treated the rest
+of the line as a comment, and `gm overlay` ran with no window argument. The
+result was a popup that flashed and vanished. The binding now resolves the window
+id in-shell with `gm overlay "$(tmux display -p "#{window_id}")"`, which works
+whether or not tmux expands the format. If you installed 0.7.0's binding, re-run
+`gm tmux install` (or, on Oh My Tmux, replace the two lines in
+`~/.tmux.conf.local`) and reload with `tmux source-file ~/.tmux.conf`.
+
+### Panes are framed, so cards read as separate
+
+Each card is now drawn inside a box border matching its pane's rectangle, so
+adjacent summaries no longer blur into one another.
+
+### The peek is instant
+
+The overlay used to block on a tmux-version check and a summary-refresh pass
+before painting anything. It now paints from cache immediately and kicks the
+background refresh off afterwards, so `ctrl-g` feels like a peek rather than a
+load.
+
+### Cards lead with the headline
+
+Each card now leads with the session's one-line headline — the scannable,
+subject-first clause — instead of the multi-sentence overview, which tended to
+open with a generic verb ("Implemented…"). Glancing across panes, the headline
+is what tells them apart.
+
+## 0.7.0
+
+**Upgrading:** nothing changes unless you opt in. Everything below is additive and
+gated behind tmux; if you don't run `gm tmux install`, `gm` behaves exactly as it
+did in 0.6.1. The overlay needs **tmux 3.2 or newer** (for `display-popup`), which
+`gm doctor` now checks for.
+
+### Peek at every agent at once, from tmux
+
+Drive your agents in tmux and `gm` can now answer "what's happening in each of
+these panes?" without you switching into any of them. `gm tmux install` writes two
+key bindings to `~/.tmux.conf`:
+
+- **ctrl-g** peeks — every pane in the current window is overlaid *in place* with
+  its summary card (what landed, what's still open, the next step, the `⚠`
+  mid-task flag). Cards paint instantly from the cache and refresh in the
+  background; any key dismisses the overlay and leaves your panes untouched.
+- **ctrl-shift-g** opens the `gm` picker in a popup, and Enter resumes your choice
+  into a new tmux window — history and live panes, one keystroke apart.
+
+The overlay maps a pane to its session by working directory and recency. For an
+exact link — including resumed sessions that share a directory — launch through
+`gm run claude` / `gm run codex resume`: it attaches your terminal as usual and
+records which pane the session runs in.
+
+`gm doctor` reports whether the overlay is available and, if not, why. tmux joins
+ripgrep and fzf as an optional companion; nothing here is required.
+
+## 0.6.1
+
+### The chat/summary split is coloured now
+
+The picker's preview pane rendered monochrome — its stdout is a pipe, so gm's
+colour gated itself off, and even the `── ask ──` divider between the session
+card and the chat came out plain. fzf paints the preview with `--ansi`, though,
+so the seam can carry colour: the divider is now **cyan** (gm's own accent) and
+the `you` / `gm` speaker labels light up, while the card stays monochrome.
+
+It's an accent, not the message — the divider's glyphs and the speakers' layout
+still carry the structure, so `NO_COLOR` and `TERM=dumb` lose only the colour.
+`gm ls` and `gm show` are unchanged and still pipe clean.
+
+## 0.6.0
+
+**Upgrading:** `ctrl-o` in the picker no longer suspends the list. It opens the
+chat in the preview pane instead, so the session you were reading stays on
+screen. Nothing else changes for you: bare `gm ask`, the `--json` form, and the
+`fzf < 0.46` and no-fzf fallbacks all behave exactly as before.
+
+### ctrl-o asks in the pane, not over the list
+
+`ctrl-o` used to suspend fzf and hand the child a full-screen REPL — so the first
+thing it did was take away the very session you pressed it *to ask about*.
+Browsing and asking are the same activity, and that made them mutually exclusive.
+
+Now `ctrl-o` is a **mode**, not a launch. The list stays put, the card moves to
+the top of the preview pane, and the answer arrives underneath it while you keep
+arrowing around. It's **one continuous thread**: moving the cursor re-points
+"this session" without forking or resetting what you've asked. `esc` leaves chat
+and hands back the list exactly where you were.
+
+`claude -p` buffers rather than streams, so there is no answer to render word by
+word — instead a **1-second heartbeat** shows `thinking… 14s` while the request
+is in flight, then the answer lands in one paint. The picker never freezes, and
+`esc` cancels a request mid-answer.
+
+Nothing regresses for people who never ask: with no conversation yet, the card
+gets the full pane exactly as today. `fzf < 0.46` keeps the full-screen `execute`
+REPL, and the no-fzf numbered list keeps its `a` key.
+
+### Fixes
+
+- The orphan-transcript sweep is now keyed on the **run**, not the transcript. A
+  picker killed between `ctrl-o` and the first question leaves a `.browseq` with
+  no `.jsonl` beside it; keyed on the transcript, the sweep could never see that
+  run, so the cache grew one file per killed picker. It now reaps every member of
+  a dead run.
+- Multi-byte output from providers is no longer mangled at chunk boundaries, and
+  callers can watch a request as it streams.
+
+## 0.5.0
+
+**Upgrading:** the first time you run `gm` in a terminal, it will ask you to
+choose a harness before doing any model work. Nothing prompts when the output
+isn't a terminal, and `GIGAMANAGE_SUMMARY_CMD` still overrides everything — so
+scripts, CI and agents are unaffected.
+
+### gm asks who to call, once, instead of assuming
+
+The first time you run `gm` in a terminal it asks which harness should do its
+model work — Claude Code, Codex, any command that reads a prompt on stdin, or
+nothing at all. Before, it assumed `claude -p` and started spending tokens in the
+background without ever mentioning it.
+
+Your answer lives in `~/.config/gigamanage/config.json`, and `gm setup` changes
+it. Choosing **nothing** is a real answer: `gm ls` and `gm show` still work on
+hard facts alone, and nothing calls a model.
+
+Nothing prompts unless there's a human at the other end. No TTY, `--json`, or an
+internal command means gm behaves exactly as it did before: autodetect and carry
+on. `GIGAMANAGE_SUMMARY_CMD` still overrides everything, so existing scripts and
+CI need no changes.
+
+### gm ask
+
+`gm ls` answers "what was I doing?" one row at a time. **`gm ask`** answers the
+question that spans them:
+
+```bash
+gm ask "what's still broken?"
+gm ask "what did I already try for the retry?" --json
+```
+
+It starts from the summaries already on disk, so a question costs one model call
+rather than a scan of your transcripts. When the summaries aren't enough it runs
+`gm grep` against the real thing and reads what you actually said.
+
+**In the picker, `ctrl-o`** opens it on the session you're highlighting and drops
+you back in the list, right where you were, when you're done. Without fzf, the
+numbered list spells it `a`.
+
+Not `shift+f`: fzf's query line eats plain letters, so `F` would just type an
+`F`. Not `alt-a` either — macOS Terminal and iTerm2 send `å`.
+
+### The picker explains its markers
+
+`gm ls` printed a key for `⚠`, `◐` and `○`. The picker — bare `gm` — rendered the
+same three markers and explained none of them, which put the explanation exactly
+where you needed it least: `ls` is the command you run to read a list, and the
+picker is the one you run to *choose*. `⚠` is the whole point of the tool, and in
+the picker it was an unexplained glyph.
+
+Both picker paths now carry a key: a second header line under fzf, and a line
+above the "install fzf" hint in the numbered fallback.
+
+It is deliberately static — every marker, always, and never a count — while
+`gm ls` keeps its counted one. fzf sets its header once, at spawn; ctrl-r
+replaces the list and leaves the header alone. Counts there would freeze at open
+and be wrong after the first refresh, which is precisely when they change. A key
+that is stale exactly when it matters is worse than no key at all.
+
+## 0.4.0
+
+### ctrl-r refreshes the picker
+
+The picker used to be a dead end: the list it opened with was the list you were
+stuck with. Sessions you started since never appeared, and rows marked `○` stayed
+`○` however long you sat there.
+
+**ctrl-r** now reloads to your most recent sessions and starts summaries for any
+that need one, without leaving the picker — so it's something you can navigate in
+while an agent works alongside you. Without fzf, the numbered list takes `r` for
+the same thing.
+
+Repeated presses are safe. The lock that already stopped five `gm ls` from
+starting five summarizers stops this too: a press while a pass is running just
+reloads. Sessions whose summary is already current are never rewritten, so ctrl-r
+on a fresh list costs nothing.
+
+### Bare `gm` summarizes what it shows
+
+Only `gm ls` kicked off a background pass; the picker never did. It does now,
+over the sessions it is about to offer, and rows being written are marked `◐`
+there as well as in `gm ls`.
+
+### Fixed: `--no-auto-summarize` never worked
+
+`gm --no-auto-summarize ls` spent tokens anyway. The flag is declared on the root
+command, and commander does not copy root options into a subcommand's own
+options — so the check read `undefined`, compared it against `false`, and
+concluded you wanted summaries. Only `GIGAMANAGE_AUTO_SUMMARIZE=0` actually
+turned them off.
+
+The flag now works, on `gm ls` and in the picker, and it is carried across to the
+process ctrl-r starts.
+
+### Shorter headlines
+
+Row headlines asked the model for "max 80 chars" and then rendered them in a
+72-char column — an overflow by construction, read as a truncated sentence. They
+are now one scannable clause, sized to the column they live in.
+
+The summary cache key covers the prompt as well as the session, so this reaches
+summaries already on disk: they regenerate in the background on first run rather
+than keeping their old headlines forever. That costs a pass of model calls once.
+`GIGAMANAGE_AUTO_SUMMARIZE=0` still opts out of all of it.
+
+## 0.3.0
+
+### Summaries keep up with what you actually look at
+
+The background pass used to cover a fixed **10** sessions while `gm ls` displayed
+**20** — so the bottom half of the default view was permanently marked "no summary
+yet", and the feature looked broken even though it was working exactly as built.
+
+The window now follows the list: `gm ls` keeps 20 summarized, `gm ls -n 50` keeps
+all fifty. Summaries are written **8 at a time** in parallel (tune with
+`GIGAMANAGE_SUMMARY_CONCURRENCY`), and a single pass writes at most 50, saying so
+rather than truncating in silence.
+
+### You can see it working
+
+Rows being summarized *right now* are marked `◐`, distinct from `○` ("no summary
+yet, nothing running"). The decision is made before the list renders, so the icon
+is true on the very run that starts the work.
+
+### Background failures are no longer silent
+
+The worker's stdio is discarded, so a broken provider used to mean summaries
+simply never appeared, with nothing to look at. Failures now land in
+`~/.cache/gigamanage/auto-summarize.log`, and `gm doctor` surfaces the last one.
+
+**Fixed:** the worker could silently write **zero** summaries. It resolved its
+queue by loading "the N most recent sessions" and filtering — but with sidechains
+included, the most recent N are mostly subagent transcripts, so the filter matched
+nothing. It now looks the queued sessions up across the whole store.
+
+### The picker wraps too
+
+fzf rows no longer truncate: a session is one NUL-delimited multi-line record
+(`--read0`), so a long summary wraps and is still selected as a single item. fzf
+below 0.46 has no multi-line display, so it falls back to single-line rows rather
+than rendering one session as several bogus entries. The numbered fallback (no fzf
+installed) wraps as well.
+
+## 0.2.0
+
+**gigamanage now spends tokens on your behalf unless you tell it not to.** That is a
+change of default behavior, which is why this is a minor bump rather than a patch —
+everything in it already shipped as 0.1.3/0.1.4, but the version number was
+under-selling it.
+
+### Summaries write themselves
+
+Any `gm` command now checks the 10 most recent sessions and, if any lack a current
+summary, writes them in a **detached background process**. The foreground command
+never waits on a model: it prints, tells you on stderr what it started, and exits.
+Summaries appear on your next run. Rows still waiting are marked `○`.
+
+Three things this deliberately does not do:
+
+- **Block.** A summary costs ~8s of model time. Ten of those inline would turn a
+  60ms `gm ls` into a minute of waiting.
+- **Stampede.** A lock in `~/.cache/gigamanage` means five `gm ls` in a row start
+  one summarizer, not five.
+- **Loop.** The summarizer *is* `claude -p`, which writes a session of its own.
+  Automated runs and sidechains are excluded from the target set, so gigamanage
+  cannot summarize its own summarizer forever.
+
+**Turning it off**, because background model calls cost money:
+
+```bash
+gm --no-auto-summarize ls          # once
+export GIGAMANAGE_AUTO_SUMMARIZE=0 # for good
+```
+
+It also stays quiet when no summary provider is installed — a missing `claude`
+never breaks a read command. `gm doctor` reports the current state.
+
+### `gm ls` wraps instead of truncating
+
+Long descriptions were cut off at 72 characters, so the sessions with the most
+informative summaries were exactly the ones you could not read. They now wrap to the
+terminal, with continuation lines indented under the description column.
+
+Piped output still emits **one line per session**, untruncated, so `gm ls | grep`
+behaves. The fzf picker's rows stay single-line, because fzf maps lines back to
+session ids.
+
+## 0.1.2
+
+- `gm --version` reported a hardcoded `0.1.0` regardless of the installed version.
+  It now reads `package.json`.
+
+## 0.1.1
+
+- First release published through [trusted publishing](https://docs.npmjs.com/trusted-publishers/)
+  (OIDC): no long-lived npm token, and every release carries a provenance
+  attestation binding the package to the commit and workflow that built it.
+
+## 0.1.0
+
+First release.
+
+- Index sessions from **Claude Code** and **Codex**; one adapter per harness, and
+  adding another is a single file.
+- Summaries describe where a session **landed**, not where it started — built from
+  the *tail* of each transcript, because a harness's own title is written in the
+  first few seconds and never revised.
+- Sessions that ended mid-task are flagged `⚠`.
+- `gm resume` hands off to the right CLI (`claude --resume` / `codex resume`) in the
+  session's original directory.
+- `--json` on every read command, so agents can call it too.
+- Read-only: never writes to a session file.

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 James Gao
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.