Changelog
quartermaster tells you how much quota is left in every agent account on the machine — every Claude Code profile, not just the default one, plus Codex and the providers stored by opencode. The number is read from disk, with no network call and no account; credentials are only checked for whether they exist and when they expire.
Install with brew install legiosai/tap/quartermaster or npm install -g @legios/quartermaster; the other channels are in the README. Every version below is a git tag and a GitHub release.
Each entry says what a user would notice. Refactors, tests and CI work are left out unless they changed what you see.
v0.1.30 — 2026-11-17
Changed
- The npm package is published through trusted publishing. Its provenance on npmjs.com now comes from this repository's release workflow over OIDC, so publishing no longer depends on a long-lived npm token. The apt repository keeps the same URL and signing key; it is now built on its own branch. The app itself is the same as 0.1.29.
v0.1.29 — 2026-11-17
Fixed
- A refresh that can't reach the server says what blocked it. Every network failure used to read
couldn't get the quota: fetch failed, whatever had failed. On a Mac it read that way for three days while an application firewall dropped qm's connections. It now names the layer and who to suspect:no DNS answer for api.anthropic.com (ENOTFOUND),the connection to api.anthropic.com was refused (ECONNREFUSED): a firewall or a proxy?,something on this machine dropped the connection to api.anthropic.com (EBADF): an application firewall?, or a certificate that doesn't verify because something is intercepting TLS. z.ai's refresh says the same, and Codex says it couldn't reach its server instead ofthe response had no rateLimits.
Added
qm --diagnosticotests the network. A new section resolves api.anthropic.com, connects to port 443 and does the TLS handshake, and says which step failed. It runs from qm's own process, because an application firewall lets one program through and blocks another:curlcan fail while qm works, and the other way round. If the check fails and Blackice, Little Snitch, LuLu or Radio Silence (macOS) or OpenSnitch (Linux) is running, the verdict names it as the likely cause and says what to allow. It sends no HTTP request and no credential, takes at most five seconds, and--sin-redskips it.
Changed
- Short plan names. The plan chip in the panels was cut off (
Ultra (max_2…). Plans now readUltra · 20x,Max · 5x,Team · 5x,Pro, and Codex'sPro LiteorPlusinstead ofproliteandplus.qm --jsoncarries them inplanNombre;plankeeps its old value. - The Windows tray panel, redesigned like the GNOME panel and the macOS menu. It was the last of the three with the old layout. It now has the same pieces:
- The header is a card tinted with the colour of the state, and it gives the reset as a local time too, plus how old the reading of that bar is.
- Each account says on top how old its number is and how often it is read, with the plan as a chip next to the name.
- The bars have names you can read (Sesión, Semanal, Semanal (Fable)). Each bar shows the time to its reset next to the percentage, and the one that stops the account is marked in the margin.
- The curve of the bar that stops you carries on, dashed, to the reset, with the stretch you'd spend without quota marked in red. A curve that didn't move says so in words instead of drawing a flat line.
- The pace, the daily budget (with its ring) and why an account is dormant sit under a rule.
- Accounts with no number share one block at the end, and commands in their message are set in monospace and never split across lines.
v0.1.28 — 2026-11-17
Changed
- The macOS menu, redesigned like the GNOME panel. The menu bar's dropdown had kept the old layout. It now has the same pieces as the full GNOME panel:
- The bars have names you can read (Sesión, Semanal, Semanal (Fable)). Each bar shows the time to its reset next to the percentage, and the one that stops the account is marked in the margin.
- The header is a card tinted with the colour of the state, and it gives the reset as a local time too.
- Each account says on top how old its number is and how often it is read, with the plan as a chip next to the name.
- The curve of the bar that stops you carries on, dashed, to the reset, with the stretch you'd spend without quota marked in red.
- The pace, the daily budget (with its ring) and why an account is dormant sit under a rule.
- Accounts with no number share one block at the end instead of one card each.
- A pace whose 100 % is already in the past no longer reads
100 % en 0s.
v0.1.27 — 2026-11-17
Fixed
- Several Claude Code sessions open on one account no longer inflate the daily budget. Each open session redraws its statusline with the quota of its own last API response, so an idle one keeps repeating an old number. With four sessions open on a Team account, the stored weekly bar flipped between 18 % and 19 % every seventy seconds, and the budget counted each return to 19 as spending:
spent 219 % todaywith the weekly at 19 %. Within one window usage never goes down, so the statusline now keeps the higher number, and ignores a reading from a window that has already ended. A real reset carries a later reset time and still goes through. The budget also stopped counting a climb back after a small drop as new spending: it measures from the highest reading since the last reset, so history recorded before this version reads right too.
v0.1.26 — 2026-11-17
Added
- The GNOME panel draws where you are heading. The curve of the bar that stops you used to end at the last reading, and the pace was a sentence under it:
3.7 pts/h · 100 % en 15h56m — antes del reinicio. The curve now carries on as a dashed line, at that same pace, up to the bar's reset. If it reaches 100 % first, the stretch from there to the reset is drawn in red and says how long you would be left without quota (sin cuota 2d5h), with the time you hit 100 % under that point and the reset at the end. If it doesn't, the line ends at the reset with the percentage you would reach (~79 % al reinicio). When the reset is days away, the past keeps a third of the width and a dotted mark at «now» shows where the time scale changes. The line is drawn only when the pace sentence is, so the two can't disagree.
v0.1.25 — 2026-11-16
Added
qm --diagnosticoends in a verdict, on every platform. It used to list what it found and conclude only about the GNOME extension. Now the last section says what is broken and the command that fixes it (or that nothing is), and a new section lists each account's number with its source and age, so «is the quota fresh?» has an answer. Accounts you hid or haven't used in weeks are listed but not flagged.qm --diagnosticoon a Mac looks at the menu bar, and on Windows at the tray. The first report from a Mac showed the GNOME section, said nothing about the menu bar, and listed the menu bar's compiled binary in the cache in a way that read like a bar that had stopped a week ago. It now says whether the menu bar is running, which login item starts it and how it last exited, with the last lines of thebrew serviceslog; on Windows, whether the tray is running, whether it starts at sign-in, and the end of its log. The cache section says what each file is.
Fixed
- The daily budget no longer counts the same points over and over. With
qm --statuslineset up, the statusline and the endpoint both report the weekly bar, and they don't always agree — by a rounding point, or (on one Team account) 4 % against 16 % for the same window. The bar on screen is whichever reading is newer, and the history recorded both, so it zig-zagged between them, and every climb back up counted as spending. One account showedspent 802.0 % todaywith the weekly at 16 %. The pace and the projection were fitted to the same zig-zag. Each reading in the history now says where it came from, and every number drawn from it uses one source only: the source of the bar you are looking at. History recorded before this version is left out, so on the day you update the pace takes a few readings to come back and «spent today» starts from the update. qm --diagnosticoread the journal again. It asked for the whole boot and filtered afterwards; on a machine up for a day that is several MB, past the output limit, and the section said «couldn't run journalctl» with journalctl working fine. It now filters on journald's side.- One cache directory when
XDG_CACHE_HOMEis set. The history and the--avisomemory followed it, but the last endpoint, Codex and statusline readings went to~/.cacheregardless, so the cache was split in two — and the diagnostic looked for the GNOME extension's heartbeat in the wrong one and called a running extension dead. Everything now goes whereXDG_CACHE_HOMEsays; a reading left in~/.cacheis still read until the next one replaces it, so updating loses no number.
v0.1.24 — 2026-11-16
Changed
- The full GNOME panel, redesigned. The bars have names you can read — Sesión, Semanal, Semanal (Fable) — instead of the API keys; the terminal keeps the keys, which is what you search a log for. Each bar is one line with the time to its reset next to the percentage, and the one that stops the account is marked in the margin, as in the terminal. The header is a card tinted with the colour of the state, and it gives the reset as a local time too (
reinicia en 3d1h · jue 22:30). Each card says on top how old its number is and how often it is read. The trend and the daily budget sit under a rule, apart from the bars, and the budget carries a small ring with how much of today's share is spent: the same meter the account shows in the top bar, which the panel never explained. A curve that didn't move is said in words instead of drawn flat, commands inside a message are set in monospace and no longer break across lines, and the panel is a bit shorter.
Added
- A minimalist look for the GNOME panel.
qm --apariencia=minimalistaswaps the full panel for one that answers only «how much is left, and when does it come back»: a ring with what stops you first, then one row per bar with its letter (ssession,wweekly,wfthe Fable weekly), a meter, the percentage and the time to the reset. Accounts with no number and dormant ones are named in one line at the bottom. The full panel stays the default: if you change nothing, nothing changes, andqm --apariencia=basicagoes back. The setting lives in~/.config/quartermaster/config.json(qmwon't overwrite that file if it can't parse it) and reaches the panel asaparienciainqm --json. Designed by knighthw in #27. - Shell completion for bash and zsh.
qm --<Tab>completes the flags, and--solo=,--ocultar=,--cuentas=and--cuenta=complete your account names, read fromqm --json --breve. The.debinstalls both files where bash-completion and zsh find them on their own, andmake instalarlinks them under~/.local/share/(zsh needs that directory in itsfpath).
v0.1.23 — 2026-11-02
Changed
- The statusline shows the weekly, not a bar nobody can name. 0.1.22 named the unnamed bar that stopped the account first (
main 44% iguana_necktie), but a number nobody can explain still covered the weekly you actually read. Bars without a name now stay out of--breveand--statuslineuntil they reach 90 %, the point where--avisowarns, and then show with their name. Below that the line shows the weekly:main 15%. Model-scoped bars keep their label (main 75% Fable), and the full table still lists every bar.
v0.1.22 — 2026-11-02
Changed
- Desktop notifications leave the screen on their own, the urgent one too. On GNOME the «stopped» / «you hit the wall before the reset» notification went out with
criticalurgency and no expiry, and GNOME keeps those on screen until you close them. Every notification is now normal urgency with an 8-second expiry: GNOME takes it down after its own few seconds (it waits for you if you are away), and it stays in the notification list. The urgent one still has the error icon. On Windows both kinds ask for the same 8 s; Windows 10 and 11 use the Accessibility setting instead (5 s out of the box). On macOS nothing changed: it was already a temporary banner, and if it stays, Script Editor's notification style is set to Persistent or Alerts in System Settings. - The statusline names the bar that stops you when it isn't the session or the weekly. On an account where Claude Code only sends the weekly (13 %) and the bar that stops you first is one the server sends without a name, the line said
main 44%, which read like a wrong weekly. It now saysmain 44% iguana_necktie, and a model-only bar says which model:main 75% Fable. The session and the general weekly stay unlabeled.
v0.1.21 — 2026-11-01
Added
qmspeaks English. The README and landing page are in English, but everythingqmprinted was Spanish. It now picks the language from the locale like any terminal program — the first non-empty ofLC_ALL,LC_MESSAGESandLANG(es*is Spanish, anything else, includingC.UTF-8, is English), or the system's regional setting on Windows.QM_IDIOMA=es|ensets it for good and--idioma=for one run. The table,--breve/--statusline,--help, errors,--esperar,--diagnostico,--desglose(tips included),--mejor, the--avisohook's message and number formatting (5.8G, not5,8G) all follow it.--jsonkeeps its text in Spanish unless a language is asked for explicitly, because the panels draw it next to their own Spanish labels; its keys never change. A Spanish speaker on a machine withLC_ALL=C.UTF-8now gets English until they setQM_IDIOMA=es.qm --statusline: a fresh number on every prompt, with no network and no credential. Claude Code stopped leaving the quota in each profile's.claude.jsonon every use — on the machine this was found on, neither of the two profiles used daily had it — so a fresh install ofqmshowed no Claude number at all and told you to "use it once", which no longer helped. What Claude Code does on every redraw is hand the statusline command arate_limitsobject with the session and weekly percentages of the last API response. Setqm --statuslineas thestatusLinecommand and it saves that per profile and prints the usual one-line summary; every otherqmthen reads it. It carries only those two bars, so it is laid over the last endpoint reading instead of replacing it, and a bar that comes from the older reading shows its own age (read 2h ago;medidoEnper window in--json). Already have a statusline script?echo "$input" | qm --statusline.- An account with no stored quota now says how to get one — set up
qm --statusline, or runqm --refrescar— instead of "use it once". qm --mejor: start in the profile with room. It prints the directory of the Claude profile with the most headroom right now: the lowest binding bar first, a projection that hits 100 % before the reset counts against it, and the emptier session breaks ties. A window that already reset is unknown, not 0 %.qm --mejor -- claudelaunches there — and launches the default profile withoutCLAUDE_CONFIG_DIR, because with the variable set Claude Code reads a different.claude.jsonand, on macOS, skips the keychain.--jsonshows the ranking and why. Disk only.qm --aviso: a warning inside Claude Code before the wall. As aUserPromptSubmithook it shows you (not the model) a message when the binding bar crosses 90 % (--umbral=N), 95 or 100, or when the projection hits 100 % before the reset within 30 minutes, and names a profile with more room and how to start it. Once per step and window; it never blocks the prompt and always exits 0. Disk only.qm --desglosenow says what to do, not just where it went. Each account opens with a short «qué hacer» block: at most three tips, each stating the measured number it rests on and ranked by how much of the weighted cost it explains — prompts of 400k or more (/compactor/clearsooner;/newon Codex), cache rewritten after a pause, an agent type that costs per request as much as the main thread when a cheaper one is in common use (aforkcarries the whole conversation), and a few sessions taking most of the week. A tip appears only when its number crosses a threshold, so an account with nothing to fix gets no generic advice. There is deliberately no "use fewer subagents" tip: measured, a subagent request costs less than a main-thread one.--jsonaddsconsejos(clave,parte,frase) andsesionesTotalper account, and the thresholds inumbrales.consejos.- The statusline also says what the next prompt will cost.
qm --statuslinenow starts with the session's context size and how long its prompt cache stays warm —ctx 182k · cache 12m · main 31/44%— read from the same input Claude Code passes it. The context turns yellow from 400k (prompts that size were 35 % of a measured week's weighted cost with 16 % of the requests), and once the cache goes cold the line says how much the next request rewrites,cold cache (+181k): the number that decides between/clearand carrying on.--breveis unchanged. - The panels say how old each bar is when it isn't as fresh as its account. Since
qm --statusline, one account can mix a session and weekly bar from seconds ago with other bars from the last endpoint reading, and the panels only showed the newest age — labelling a 44 % bar from half an hour ago as current. In the GNOME panel, the macOS menu bar, the Windows tray andqm-web, such a bar now readsreinicia en 3d7h · leída hace 24m(· viejaafter 6 h), and so does the "what stops you first" header when it shows one. The stale mark in the top-bar item follows the bar it draws. qm --desglose: where the quota went, not just how much is left. It reads the same transcripts as the local-usage line and breaks the last 7 days (--dias=N) down per account: cost by token class, by prompt size, cache rebuilds by idle gap and by main thread vs subagent, subagents by agent type, and the most expensive sessions with their context sizes, compactions and duration. Tokens are weighted by relative API price (cache read 0.1, cache write 1.25 or 2 by TTL, input 1, output 5) as a proxy for quota drain — Anthropic does not publish how a subscription weighs them, and the output says so.--jsonhas the same numbers;--redactadoalso hides project directories and session ids. Disk only, no network: about two seconds for a week of two busy profiles (~2 GB). Codex is included; opencode and Antigravity are not, because they do not keep per-request cache usage.
Changed
- Dormant accounts fold into one line each at the bottom of the table. Each used to take five lines to say there was nothing there — on the machine this was written on, three of six accounts and fifteen of thirty-five lines. The reason each one is dormant is still printed;
qm --todasshows them in full, and if every account is dormant they are shown in full anyway. - A bar the server doesn't name reads
other limit (iguana_necktie)(otro límite (…)in Spanish) instead of the server's internal key alone. The key stays, because it is the only thing that tells two such bars apart. The panels get the same name.
Fixed
qm --jsoncame out cut off when read through a pipe, and the panels went blank. Linux gives each new pipe 64 KB, but only 8 KB once a user already has many open (fs.pipe-user-pages-soft) — a session with a lot of terminals and agents gets there.qmwrote the first 8 KB and exited before the rest: 16.6 KB of JSON arrived as 8 KB, and the GNOME indicator logged «qm no devolvió JSON» several times an hour. Every exit now waits for stdout and stderr to drain. A test shrinks the pipe to 4 KB on purpose so it fails without the fix.- A bar scoped to a surface lost its name. The usage endpoint describes a bar's scope as a model or a surface, each with a display name; the surface one was read as plain text, so it never matched and the bar showed only its internal key.
- A Claude Max 20x or Ultra account showed up as
claude_max. For an individual account the plan came fromorganizationType, which isclaude_maxfor Max 5x, Max 20x and Ultra alike, so upgrading changed nothing on screen. The terminal, the panels and the dashboard now show the plan name Claude Code shows, with the rate-limit multiplier when there is one:Ultra (max_20x),team_tier_1 (max_5x).qm --jsonadds it asplanNombre;plankeeps its old value.
v0.1.20 — 2026-09-27
Fixed
- An account with no 5-hour session showed nothing in the bar. The number and the meter of each account in the top bar were the 5-hour session, read as-is. Some plans don't have one — Codex Pro reports only a weekly window — so that account was drawn with an empty track and no number, as if it had no quota, while its weekly was the one about to stop it. The bar now shows each account's shortest window: the session when there is one, the weekly when there isn't. Same fix in the GNOME item, the macOS menu bar and the Windows tray.
v0.1.19 — 2026-09-22
Fixed
- The consumption figure 0.1.18 added to the panel never actually appeared. The change was real and the rendering was right — the data never got there. The GNOME indicator asked
qmfor--json --breve, and--breveis the mode that skips reading transcripts: it is what makes it instant, which is why it belongs in a statusline that redraws on every prompt. It also means the local figure is absent, so the panel drew every no-quota account without one, exactly as before. It now asks for the full JSON. Measured cost: 2.3 s against 0.58 s, paid once every 300 s, with burst redraws already coalesced by the debounce — 0.8 % duty cycle so the panel stops lying by omission.
Changed
- The Antigravity row now names your actual plan:
Google AI Pro, not a generic phrase. It is read from the IDE's ownstate.vscdb— Antigravity is a VS Code fork and keeps state where VS Code does. The line went from "plan de Google: el porcentaje no está en esta máquina" to "Google AI Pro: Antigravity no deja el porcentaje en disco", which is the difference between not knowing and not having looked.
There is still no percentage, and that is not for lack of looking. Searched: all 120 keys in the state database, the Local Storage leveldb, antigravity_state.pbtxt, and the IDE logs, where the only mention is literally quota undefined. The closest thing is a modelCredits entry holding two sentinels — availableCredits at 0 and a minimum of 50 — that do not move with use. Google does not publish a remaining-quota figure locally, so the account keeps reporting sin-cuota-legible and keeps showing the one thing that is knowable: what it spent.
The adapter never reads antigravityAuthStatus, which holds the OAuth access token in plain text. It is not needed for any of this.
v0.1.18 — 2026-09-22
Changed
- The Gemini account is now Antigravity, because that is the one being used. 0.1.17 read Gemini CLI's transcripts. It worked — and it measured the wrong tool: on the machine this was written on, the last Gemini CLI turn that spent a token is from 14 June, while Antigravity had spent 6.1M tokens across 560 requests that same week. The panel said
0 en 7dthe whole time. Measuring the tool you don't use is worse than measuring nothing: it answers, and the answer is false. Antigravity shares~/.geminibut does not write totmp/— it keeps one SQLite database per conversation under~/.gemini/antigravity/conversations/, with one protobuf per request.--sin-antigravityleaves the row out;--sin-geminiis gone with the adapter it belonged to. - Accounts without a quota now show what they did spend. A no-quota account used to look identical whether it had burned six million tokens this week or had not been touched since June — the terminal always printed the local figure and the panel threw it away. Not having a percentage is not the same as not having data, and the local floor is read off the disk with no network and no credential. The number sits next to the name, and only when there is something to count: a
0beside every dormant account is noise shaped like data. Still no meter, in the top bar or anywhere else — there is no percentage to draw, and an empty meter would read as "0%, all good".
Fixed
- A credential file that does not exist was reported as "unreadable". Same bug as the one 0.1.17 fixed for the Claude rows, in the adapter that replaced it; it is now written once and shared.
v0.1.17 — 2026-09-22
Added
- Gemini CLI is now one more account, with the tokens it actually spent. It had been ruled out on the grounds that nothing under
~/.geminicounted tokens — but that was a search error: only the root and~/.gemini/history(which holds just.project_root) had been looked at. The real transcripts live in~/.gemini/tmp/<project>/chats/session-*.jsonl, and every assistant turn records the exact breakdown. Measured on one machine: 10.7M tokens across 271 requests, matching an independent count of the same files exactly. Each turn is written to the file twice, so turns are deduplicated by id; both copies carry identical counts, checked against every turn on disk. No percentage, on purpose. Google's personal OAuth plan publishes no quota and no reset — not on disk, not from an endpoint — so the account readssin-cuota-legible, the same as opencode's API-key plans. That means it shows in the panel's no-quota block and not as a meter in the top bar: an empty meter would read as "0%, all good", which is a different claim and a false one.--sin-geminileaves the row out.
Fixed
- A credential file that does not exist was reported as "unreadable", and one with no expiry date as "valid (expired)". The first said the file is there and corrupt, sending you to look for a broken file that was never written. The second contradicts itself in four words. Both came from one line that had been copied twice and was about to be copied a third time: it checked why it failed before checking whether it is there, and turned a missing expiry date into a negative duration that renders as "expired". The Claude rows had the second bug already; they are fixed too.
v0.1.16 — 2026-09-17
Fixed
- The GNOME top-bar item showed the real numbers only while the mouse was over it. Half of every refresh left a stale drawing on the bar, and hovering was what made the true figure appear — so the quota looked like it changed on its own as you moved the mouse across the panel. Measured on one machine (GNOME Shell 48.7): the item read
15 | 15 | 37 | 0whileestado.jsonat that same moment said main 43, teams 21, codex 66, glm 0; seven frames into the hover it flipped to43 | 21 | 66 | 0and stayed. The old drawing was in no file on disk — it was the texture the shell had cached for that path. The item alternated its PNG between two names to dodge that cache, which does not dodge it: one refresh in two lands back on a path the cache already holds, and the shell only drops the old entry once its file monitor reports the change, which arrives after the extension has already painted. Each drawing now gets a name that is never reused, so every refresh is a path the shell has never seen. The panel behind the item had the same bug — the one you opened could be minutes old — and got the same fix.
v0.1.15 — 2026-09-16
Fixed
- The macOS menu bar could stop refreshing for good, with the process alive and its watchdog "re-arming" every minute. Measured on one machine: the watchdog logged
lo rearmoonce at 13:57 and then nothing for twenty hours, while the token had been live since 09:32 the next morning and the bar kept drawing a day-old13%. The cause was the watchdog itself: once tripped, it forced a new five-minute timer every sixty seconds, invalidating the one it had just armed, so the timer never fired. The forced re-arm now only replaces a timer whose date has already passed; a pending one is left alone. - The terminal said
13% reinicia en vencidofor a session window that had already reset. A window whose reset time has passed carries the number of the previous window; the GNOME indicator and the Windows tray already said so, the terminal was the last screen still printing it as live. Now the bar and the figure of an expired window are dimmed, its severity flag is dropped, and the footer readsreinició hace 1d9h. Inqm --brevethe expired figure becomes?(teams ~?/26%): the real number after a reset is not on this machine, and a stale13looked like it was.
v0.1.14 — 2026-09-15
Fixed
- The GNOME indicator could stop asking the endpoint and keep drawing as if nothing had happened. Measured on one machine: seven and a half hours with no refresh, the process alive, the bar redrawing
~0/30%while the server said7/31%. The cause was two timers tied at 300 s — every poll pushed the refresh deadline further out, just before it was due, so the refresh never fired. A deadline can no longer be moved further away than the one already set, and a watchdog that depends on neither loop notices a stale reading, logscalentar.paradoonce and re-arms. The macOS menu bar has the same shape and got the same two fixes; the Windows tray, which could not stall the same way, got its tick wrapped so one exception can no longer take the whole tray down. - Installed from npm, every
qmrun printed two NodeExperimentalWarninglines on stderr. That is the channelqm --calentaruses to say why an account was not refreshed, so the real reason ended up buried between them in the indicator's log. 0.1.12 silenced them for the shell launcher, which is the.deband the clone; npm runs a different entry point and was still affected. Measured: 416 bytes of stderr before, 246 after. A real deprecation warning still gets through.
Added
- On Windows, the first
qmin a terminal now offers to put the tray icon in the notification area and start it at login — the question macOS and GNOME already asked. Nothing was missing from the package:qm-trayshipped all along, and nobody was offering it.qm-tray --instalar-arranquedoes it by hand.
v0.1.13 — 2026-09-14
Fixed
- An account nobody uses could take over every screen. A Codex account on the free plan, unused for weeks and stuck at 100 %, headed the summary on all six surfaces, coloured the icon, fired notifications and carried the
!in the statusline. Worse, it owned the blocking decisions:qm --esperarslept 14 d 23 h on it, soqm --esperar && codex exec …would not have run the right half for fifteen days, andqm --umbral=80exited 3. An account now needs both evidence of no use and nothing paid waiting on the other side before it sorts to the bottom — an exhausted account you do use is exactly what this tool exists to point at. Nothing is hidden: the row stays, in full, and says why it dropped. - On GNOME 43 (Debian 12) the shell extension did not load at all, so the meter shrank into a 16 px square and the click fell through to the GTK menu that closes the moment you touch it — the two things the extension exists to prevent. It is written against the API GNOME introduced in 45; a second copy with the pre-45 wrapper now ships alongside it, and installing picks between them by
gnome-shell --version. - The Linux launcher only looked for Node in nvm, though both READMEs promise nvm, fnm and volta. It only showed from GNOME's autostart, which has no interactive shell PATH: the indicator started, found no Node, and went on showing the previous session's numbers without saying so.
qm-indicatorwas not on anybody's PATH, though both READMEs documentqm-indicator --instalar-extensionand--instalar-arranqueas commands. npm registeredqm,qm-webandqm-barra(macOS) and not it; the.deblinked onlyqm. npm now registers it too, and the.deblinksqm,qm-webandqm-indicator.- The z.ai / GLM rows said the percentage did not exist on this machine. It did.
GET https://api.z.ai/api/monitor/usage/quota/limitanswers with both windows, percentage included. Those rows now get a real bar after you ask for it once (--refrescaror--calentar), cached like every other number, so the panels show it without ever touching the key themselves. This makes z.ai the second host quartermaster ever talks to; both are listed inSECURITY.md. - The "free in" time for an opencode provider was up to 8 hours wrong, always in the direction of believing you were free early: the reset time z.ai reports in a
429is Beijing time and was being read as local. It only matched on a machine in UTC+8, which is where it was first measured. - A cached endpoint reading was stamped "just now" no matter how old it was. It showed least on the Claude rows, where the
.claude.jsoncache usually wins on date; the opencode rows have no other source, so a three-day-old number announced itself as fresh. The age now comes from when it was measured, with the same "old" mark at 6 h. - Windows tray, from several reports: every tick in "Icons in the tray" closed the panel, so choosing icons meant reopening it once per account; the reopened submenu came back in the top-left corner of the screen, loose from the panel; the panel was not centred inside its menu (8 px of air on the left, 28 on the right); and the gap between cards sat only below each one, so the panel started flush against the top edge.
Changed
- The Windows tray draws each account in a rounded card, like GNOME and macOS, and the four options at the bottom moved behind a
⋯row. The cost is one extra click for "Refresh now", which is the exception — the panel refreshes itself every 30 s.
v0.1.12 — 2026-09-14
Fixed
- On Linux there was no way to get the top-bar item unless you had cloned the repo. It was started with
make indicadorandmake autostart, and whoever installed from apt or npm has no Makefile. The firstqmin a terminal now asks the same question macOS has asked since 0.1.7, and saying yes also installs the GNOME extension — the part that makes a click open the panel. This can need a log-out: GNOME does not load a freshly installed extension on Wayland, and when that happens the closing line says so. The question only appears when there is somewhere to show the item (graphical session,python3-giwithAyatanaAppIndicator3, and a host for it);qm --diagnosticosays which piece is missing. - The npm package did not ship the GNOME extension at all. Installing that way left the item hosted by AppIndicator, where the left click falls into the GTK menu that does not hold its grab — seen from the user's side as "I touch the bar and I get no quota panel". Measured with
npm pack --dry-run: zero files fromextension/. - A
qm --calentarthat exited 1 left no line in the journal. GLib does not raise on a non-zero exit status, so a refresh that failed passed in silence — which is exactly how 0.1.9's broken refresh went unnoticed on Linux while macOS saw it in its log. The indicator now reads the exit status, records each per-account reason once (no credential, expired token) and logs acalentar.volviowhen it recovers, plus a start-up line saying whichqmit found and with which theme. - Every
qmrun wrote two NodeExperimentalWarninglines on stderr — the same channel it uses to explain things. Only that class is silenced; a real deprecation still comes through. (The npm entry point kept the problem until 0.1.14.) - The Windows tray said "0.0 % left" where the other four surfaces said 0.3. PowerShell picked the integer overload of
[math]::Maxand floored the balance, so a whole daily budget under 1 % disappeared.
Changed
- winget: every release opened its own "New package" pull request, and six were queued behind the same volunteer moderator. Releases now push to the one that is open instead of adding another.
v0.1.11 — 2026-09-14
Fixed
- The daily budget subtracted what you had spent from a figure that already had it discounted, so every point counted twice. On screen: "you can spend 14.3 %/day · spent 14.0 % since 09:27 · 0.3 % left", where the right answer was 16.1 and 2.1. Spending exactly your budget declared you over it. The core now fixes the daily allowance at the reading the measurement starts from and ships the subtraction already done; all five surfaces read it instead of doing it themselves.
v0.1.10 — 2026-09-14
Fixed
- 0.1.9 kept asking the endpoint until it answered
429: 36 failed refreshes between 12:21 and 14:43, one every 124 seconds. The 60-second floor was per request, not per account, so two accounts were dragged along by a third that set the pace. An account read less than 5 minutes ago is not asked again unless a window reset in between, and a429now stops that account for as long asRetry-Aftersays (between 10 minutes and 1 hour).--refrescarrespects it too, and says until when. - The four drawing surfaces still carried the budget phrase 0.1.7 had already fixed in the CLI — "today you have X % left", a number that went down on its own with the clock whether or not you spent anything.
- A laptop that is off at night was treated as the exception. The daily figure needed history back to midnight, which on a machine that boots at nine is missing every single day. It now reports what was actually measured with the start time next to it ("spent 13.0 % since 08:54") and says which of the two it is, rather than reporting nothing. A reading from three days ago no longer counts as the day's baseline.
- On Windows, "today" ended at 21:00. The tray runs
qminside WSL, and a fresh WSL starts in UTC. The machine's own IANA time zone is passed through now — it has to be the IANA name, because Node silently ignores the old POSIX form and stays in UTC.
Changed
- The Windows tray icon's meter is the session, not the weekly bar. At 16 px the weekly one moves about two pixels a day and sits near half all week whatever you do. The icon has no text label next to it, so that meter is all you see at a glance — and what you want at a glance is whether you can keep going now. The weekly bar and the daily budget are still in the panel and the tooltip.
v0.1.9 — 2026-09-14
Fixed
- The macOS menu bar was refreshing the wrong accounts, and then none at all. It still used the rule GNOME had replaced the day before, so it ignored the account you were actually using if it was low, ignored one whose window had already reset, and polled a bar pinned at 100 % every minute. It also sent short account names where
--cuentascompares full ones, so a filtered refresh asked no Claude account whatsoever: with 0.1.8 installed,--cuentas=personalexited 1 with "no account to ask".--cuentasnow accepts both forms, for bars already installed.
v0.1.8 — 2026-09-14
Fixed
- The macOS menu bar's log threw away the only lines worth reading. Refresh exit codes went to
/dev/null, a failed read showed only in the menu, a notificationosascriptnever displayed left no trace, and no line had a timestamp. Lines are timestamped now, the file is trimmed past 1 MB, failures and recoveries are both recorded, andqm --calentarsays on stderr why each account was not refreshed.
v0.1.7 — 2026-09-14
Fixed
- Coming back from suspend left old numbers on screen, marked "expired". Measured after 4.1 hours asleep: the drawing loop alive and the quota 5.2 hours stale, all three sessions marked expired with their reset 1.7 hours in the past. Three separate causes — a refresh already in flight killed the timer source for the rest of the process's life, a stuck flag had nothing to release it, and GLib's timers run on the monotonic clock, which does not advance while the machine sleeps, so four hours asleep are zero seconds to the timer. The indicator now listens for logind's
PrepareForSleepand asks on resume. - The bar was following the account you were not using. Refresh picked accounts by level, which is right for an idle account and exactly backwards for the one you are spending: with two sessions open, a teams account at 2 % was asked once in 55 minutes while everything went to a Codex account at 85 %. Accounts used in the last ten minutes are now asked as well, wherever they sit. (The field that carries "last used" was being nulled by
--breve, which is the mode the bar runs — fixed in the same release.) - "today you have X % left" promised a subtraction it never did. It was the even split of the hours remaining in the day, so with the quota untouched it went from 10.5 % at 01:00 to 0.5 % at 23:00 — telling you at eleven at night that you had 0.5 % left of a daily budget you had not spent. It now subtracts what actually went up since midnight, counting only upward jumps so a window resetting mid-day cannot produce a negative, and reports nothing (with a phrase saying why) when the history does not reach the start of the day.
Added
- The first
qmin a terminal on macOS asks, once, whether to put the meter in the menu bar and start it at login. No install path could do it on its own: brew runs the install in a sandbox that cannot write to~/Library, and the npm postinstall stopped registering anything in 0.1.6. It never asks under--json,--breve,--watchor--umbral, without a terminal, or in CI. Without the question,brew services start quartermaster.
v0.1.6 — 2026-09-13
Fixed
- The GNOME panel closed the instant you touched it, and nothing said why. The shell extension was failing to load with a type error, and without the extension the panel falls back to the GTK menu — the one that cannot hold its grab and closes as soon as you touch it. A load error at start-up looked, from the user's side, exactly like a panel that closes on its own.
- The account that was blocked was the only one never asked. Refresh skipped anything at 100 %, a rule that is correct about how often to ask and had leaked into who to ask — so the one account waiting to be freed was excluded from every refresh, and "you're free again" is the most useful thing this program says. Measured: the cache said 96 % and 100 %, both critical, with the reset two hours past; a single refresh returned 2 % and 0 %, both normal. Nine hours telling someone they were blocked while they were free.
- A reading from a window that has already closed is now marked expired rather than shown as an old measurement of the same thing. A closed window's
criticalno longer paints the tray red for something that is over. No invented 0 %: after a reset the real number is unknown until it is read. - On Windows, the npm package installed and then refused to run. npm builds its
.cmdshim from the shebang, and#!/bin/shproduced one that invokes/bin/sh, a POSIX absolute path: exit 1 with "The system cannot find the path specified". It was first measured on a Windows 10 VM with no Git, but the condition is not a missingsh— GitHub'swindows-latestrunner, which hassh.exeon PATH, failed the same way.npm run construirhad also never worked on Windows, which matters because the npm tarball, the installer and CI all hang off that build. qm --versionanswered "unknown option" with exit code 2, which is the first thing anyone types. Help and version exit 0; only an option that does not exist exits 2.- On Windows, a missing credential is no longer hedged. The message used to say Windows was unverified and might be using DPAPI, which sent people off to log in again for no reason. Reading Claude Code's own binary settled it: the credential store has no per-platform branch, and DPAPI, CredRead, wincred and
safeStorageappear zero times in 206 MB. The most common Windows case — someone who simply has not logged in — now gets the right diagnostic.
Added
qm --diagnostico: one command whose output you can paste straight into a public issue — home paths become~, and no e-mail, token or account id is ever printed. It runs the five commands nobody should have to know and states the conclusion instead of leaving you the data, including the case that looked like "the fix didn't work": GNOME is still running the old extension module and the only step left is to log out. Anything it cannot find out, it says so; it works with none of those commands installed.- The GNOME indicator writes what it does to stderr, which systemd's journal picks up (
journalctl --user -b | grep qm-indicator).vivio_msis the one that answers "the panel closes on its own": a panel that lived 40 ms was not closed by you.QM_SILENCIO=1turns it off. qm --jsondeclares its contract:esquema: 1and the package version. That JSON is read by four surfaces, the statusline, the waybar module and anyone writing their own, and without a number the only way a consumer learns something changed is that their screen breaks.docs/esquema-json.mddocuments it field by field.- Windows without WSL:
qm.cmd, a per-user installer with no UAC (qmon your PATH, the tray in the Start menu), and generated manifests for winget, scoop and the AUR, plus a Nix flake and a waybar module. SECURITY.md, with the table of what is read and what leaves the machine, and the undocumented endpoint stated plainly rather than waiting to be asked.- The GNOME Shell extension was submitted to extensions.gnome.org.
Changed
- The npm postinstall no longer installs a macOS LaunchAgent. A package that reads OAuth credentials should not register login agents nobody asked for, and the path was not reliable anyway —
npm ci --ignore-scriptsis normal. Upgrading does not remove an agent 0.1.5 wrote: delete~/Library/LaunchAgents/com.legios.quartermaster.barra.plistif you do not want it. The explicit replacement isqm-barra --instalar-arranque, and from 0.1.7 the first run asks. - The README is now English and short. The 1711-line version — the engineering log, with the three panel architectures and the assumption about the macOS menu bar that turned out to be false — moved to
BITACORA.mduntouched.LEEME.mdis the Spanish README. The landing page is now in English too, with the Spanish one under/es/. - One tag publishes everywhere. The release workflow builds every artifact and pushes to each channel that has its secret configured, skipping the others by name and saying so, so a catalogue can no longer sit on an old version because somebody forgot a step.
v0.1.5 — 2026-09-11
There is no v0.1.4 tag. 0.1.4 reached npm from a tree that was missing six commits — it went out without the daily budget or the new tray. 0.1.5 is the version that carries them.
Fixed
npm install -g @legios/quartermasterinstalled and then failed on every run. Node refuses to type-strip files undernode_modules, which is exactly where a global install puts the package, soqmdied with an internal Node error each time it was invoked. The tarball now carries a compileddist/; a clone,make instalar, the.deband brew still have no build step.- The GNOME panel closed as soon as you touched it to scroll. In a GTK menu, releasing the button over an item activates it, and activating closes the menu. Underneath that was a harder wall: a GTK menu on XWayland cannot hold its grab against real desktop activity — measured lifetimes of 70 ms to 4 s with somebody using the machine. The panel moved inside the GNOME extension, where it is a compositor actor and scrolls and closes like any other shell menu. The GTK menu stays as the fallback for people without the extension.
- The apt instructions on the documentation page pointed at a file that does not exist, and had since the page was published. Copying the block left an empty keyring and
apt updatefailed withNO_PUBKEY, which says nothing about the download being the problem. The repository itself was fine.
Added
- "How much can I spend today", which is not "when do I crash". A line that divides what is left of the long window by the days until it resets:
you can spend X %/day · Y % left today. It deliberately does not split a 5-hour window, which resets four times a day, and says why instead of printing a number when it is less than an hour from the reset. Inqm, the dashboard, the GNOME panel, the Windows tray and the macOS menu bar. qmruns natively on Windows, measured with Windows' own Node rather than WSL's: it findsC:\Users\<user>\.claudeand.codexwith their backslash paths and reads Codex's quota from disk, no network and no credential.--breveworks too, so the statusline does.- Windows tray, largely rebuilt: one general icon plus one per account, each opening its own panel; dark mode; the "you're free again" notification; notifications for every bar rather than only the one that stops you first; a warning when Windows hides your icons, because a hidden icon and one that never started look identical; a per-account tick list to choose which icons appear, editable from the panel itself; and a guard so an interface exception reports itself instead of killing the tray.
- A documentation page, with every flag, the
qm --jsoncontract field by field, and when the rate line is deliberately not drawn. The site moved fromlegiosai.github.io/quartermastertoquartermaster.legios.com.ar; the old URLs redirect.
v0.1.3 — 2026-09-10
Fixed
- Codex showed a
plusplan on a machine with no subscription. It looked for the newest session file that had bars, and found one from May — reporting a plan that had been cancelled and a percentage 127 days old as if it were current, with bars drawn for an account that pays per use. The authoritative signal (auth_mode) is now read first, and an API-key account says so instead. - The opencode rows appeared on no screen at all. They were skipped under
--breve, and every surface asks for--breve: the macOS bar, the Windows tray, the GNOME panel and the statusline. The row only existed if you ranqmby hand in a terminal. A z.ai key used the day before — 56 325 tokens in seven days — was invisible, which is the bug this project exists to fix, with a different vendor. The cost that justified the exclusion was one slow query; fixed, all of opencode takes about 25 ms. - The GNOME panel said "cache from 3067h28m ago" where Windows said "127d19h". The Python duration ladder had no days branch. All four implementations were then aligned on the CLI's convention, with a gate that runs the four against one frozen table — three of them had agreed with each other, so every comparison between screens had looked reassuringly fine.
- Long account names drew outside their card and off the panel. Anything carrying user data is truncated now, and the name gives way before the number.
- Two instances left two items in the bar, opened two panels on one click and overwrote each other's PNG — trivially provoked by autostart plus a manual start.
- If GNOME crashed without shutting the extension down cleanly, the indicator hid its own item forever — no icon anywhere, which is the worst possible ending. The extension now refreshes a heartbeat and the indicator only believes it if it is recent; measured, the icon comes back on its own in 45 seconds.
- Switching the desktop to a light theme left the number in the bar white on a light background until you restarted the indicator.
Added
- A signed apt repository, so on Debian and Ubuntu
apt upgradebrings new versions by itself and apt verifies the signature before installing. - The GNOME indicator updates itself.
apt upgradereplaces the files and leaves, but the running process keeps the old code in memory — and nobody restarts the indicator, because the indicator is precisely the thing you leave running and forget. It now watches its own source and relaunches, compiling the new version before jumping and keeping the working one if it is broken. - The Windows tray draws the full panel — header ring with what stops you first, and the rate curve next to the bar that stops you — with one tray icon per account.
- The landing page, the logo, and screenshots taken from a redacted fixture, so there is no e-mail or home path in a public image.
v0.1.2 — 2026-09-10
Fixed
- The GNOME panel stopped opening on any screen it did not fit in — which is any 768 px laptop with three accounts. A GTK menu scrolls item by item, and the whole panel was a single item: with nothing to scroll, the menu simply does not open. The panel had been right at the edge of the usable height and further work pushed it to 946 px. It is one item per card now.
- After the first click, the menu never opened again. Removing a widget from a GTK container destroys it, so keeping the separator and the "Refresh" entry to reuse them left two dead objects behind. Verified with six clicks: the first opened, and then nothing.
Added
- Three ways to install without cloning the repository: Homebrew (
legiosai/homebrew-tap), npm, and a.debattached to the release. None of them compiles anything — all you need is a Node 22.6 or newer.
v0.1.1 — 2026-09-10
Fixed
- On a machine with no accounts yet,
qm --jsonprinted a Spanish sentence instead of JSON, so any statusline or script parsing it broke on the very first run — which is the first thing anyone who installs this sees.--jsonis a contract: it returns the envelope withperfiles: []. It also no longer exits 1, because having no accounts is an answer, not a failure.
v0.1.0 — 2026-09-10
First public release. The starting point was a monitor that polled every 120 seconds for hours, logged credential expired — gating poll, and displayed nothing at all, because it read one credential — the default directory's — on a machine where the two real accounts lived somewhere else.
Added
- Every account, not one. Every Claude Code profile (
CLAUDE_CONFIG_DIRand all), plus Codex and the providers stored by opencode — GLM, MiniMax, Kimi and Grok, read from its session database. - The number needs no network and no credential. Claude Code stores the last quota response it received in each profile's
.claude.json, so quartermaster reads it from disk — which means it still works on a profile whose token expired, the exact state the motivating monitor died in. The endpoint is a refresh (--refrescar), never the only path to a number. - Every bar the server sends, not the two with famous names. On the account this was measured against,
five_hoursaid 8 % andseven_daysaid 59 % while the bar actually about to stop the account was aweekly_scopedone at 75 % with severitywarning, reachable only throughlimits[]. Reading the well-known keys is not a simplification, it is a 67-point error that looks calm. - A projection, and notifications. The quota is a snapshot, so the readings are kept and a rate is fitted to them: points per hour, when you hit 100 %, and whether that lands before the window resets — with explicit guards and a stated reason when it refuses to draw a line. Notifications at 80 % and 95 %, when the projection starts saying you will crash, and when a window resets after you had been against the ceiling, which is the only one that lets you do something different right now.
- Five places to see it:
qmin the terminal (qm --brevefor a statusline, ~85 ms because it skips transcripts), the GNOME top bar with its own Shell extension, the macOS menu bar, the Windows notification area, and a local dashboard on127.0.0.1. --umbral=N(exit 3 over the threshold),--esperarto block until the quota drops so you can chain a command behind it,--watch,--redactado, and--solo=/--ocultar=plus a config file to choose which accounts show.- The rules live in one place.
qm --jsonemits the resolved answer —mostrar,frena,sesion,semanal— and the renderers read fields. The rule that stops a 75 % warning bar hiding behind a calm 8 % had ended up copied into Python, JavaScript and Swift, which means three answers the day somebody edits one. - A 60-second floor on anything that leaves the machine. An adaptive cadence had dropped to 20 s per account, roughly 360 requests an hour against an undocumented endpoint, for reading your own quota. Reading the disk stays free and has no floor.