The complete catalogue of Claude Code Tray's own CLI flags — every preview, capture, headless read-out and fixture the .exe exposes, with what each one is for. Use when you need a flag that AGENTS.md's short list does not name, when you are unsure which entry point shows a surface, or when you add a flag and need to know where it is written down.
Scanned 9/5/2026
Install to Claude Code
Skills Directorynpx -y skills add alegauss/claude-tray --skill dev-flags --agent claude-codeInstalls into .claude/skills of the current project.
Are you the author of Dev Flags?
Add the live security badge to your README — it updates automatically with every re-scan.
[](https://www.skillsdirectory.com/skills/alegauss-dev-flags)More formats (shields.io, HTML) on the badges page.
---
name: dev-flags
description: The complete catalogue of Claude Code Tray's own CLI flags — every preview, capture, headless read-out and fixture the .exe exposes, with what each one is for. Use when you need a flag that AGENTS.md's short list does not name, when you are unsure which entry point shows a surface, or when you add a flag and need to know where it is written down.
---
# Dev flags (Claude Code Tray)
Every flag the app's own `.exe` exposes. **This file is a reference, consulted rather than read** —
which is why it lives here and not in [AGENTS.md](../../../AGENTS.md), whose whole content is loaded
on every turn against a line budget (T191). AGENTS.md keeps the handful of flags that carry a *rule*;
everything else is here, complete.
Run any of them with `dotnet run -- <flag>` from the repo root, or against a built binary at
`bin\Debug\net10.0-windows\win-x64\ClaudeTray.exe`. With no flag at all the tray app itself starts.
## Build and publish
```
dotnet build -c Debug # fast compile check
dotnet run -c Release # build + run the tray app
dotnet publish -c Release # single self-contained .exe -> bin\Release\net10.0-windows\win-x64\publish\
```
## The windows, as previews
Three different hosts, and the difference matters — see UI convention 7 in AGENTS.md.
```
--main [dest] # the WHOLE window as the tray opens it: nav strip + destination
# Statistics | Context | Settings, matched case-insensitively.
# A name that is none of them prints the three and exits 1
# (T262); naming none opens Statistics, which is not a typo.
# The same refusal guards the [page] of --settings, --settings-
# tray and --capture-settings, out of the six the sidebar has.
# (Statistics | Context | Settings), under the WinForms pump the
# tray uses. Use this to look at the shell.
--settings [page] # just the Settings page, no nav strip, under a WPF pump. Any page
# name works: General, Display, "Claude Code", Notifications,
# System, About.
--settings System --sample [--reveal] # ...over the synthetic AccountFixture profiles instead of this
# machine's, unmasked with --reveal. Any published shot of that
# page must use it.
--settings System --sample week=<name># which stored week the personal fixture profile gets, and so which
# branch of the extra-usage row renders (T200): spending (the
# default, "in use now (42%)") | zero ("not in use") | absent (no
# reading, so just "Enabled") | refused ("Not available" plus the
# reason the API gave, T224 — the branch no local file can produce,
# so it writes a header-probe log instead of a usage history). An
# unknown name is refused with the
# catalogue, and **--sample that cannot be honoured stops the run**
# rather than falling back to this machine's real account.
--settings-tray [page] # ...hosted the way the TRAY hosts it (WinForms pump). The only
# preview that can see a keyboard bug.
# `ghost-pieces` is the one row whose two halves disagree ON
# PURPOSE (T301): hours the header called over, above a ghost
# line peaking at 80% because the fold lost a delta at a reset.
# It is the only way to LOOK at what T299 draws - the pinned
# mark and the floor sentence, which is on the ghost's hover, so
# `--stats ghost-pieces` and not a capture is what shows it.
--stats [variant] [modifiers] # just the Statistics page on a synthetic reading. The variants and
# modifiers are one table, src\Cli\StatsPreviews.cs, which
# --capture-stats reads too; pass a name it does not know and it
# prints the whole catalogue and exits 1 (T186).
--stats [variant] --sample # ...with the profile picker filled from AccountFixture instead of
# this machine, which is what --capture-stats --sample publishes.
--context --window [slug|name|all] # just the Context Load page, optionally opened on one project
--context --window --sample --lang en # ...over the bundled fixture, in English (for screenshots)
--context --window <slug> --scroll # ...scrolled to the source table
--context --window <slug> --simulate # ...with the 3 heaviest sources ticked (the what-if)
--context --window <slug> --demo-history # ...with a synthetic drift series behind it
--tooltip [variant] # the TRAY TOOLTIP's composed text, printed with its char count
# against the 127 cap (T214). No variant prints them all: ok
# (the default) | track | danger | remaining | profile | extra
# | extraunspent | billingelsewhere | billingremaining | atlimit |
# stoppedelsewhere | connecting | signedout | error. The first two
# billing rows differ by whether a
# cent has been spent, which is what decides where the news is
# written (T222); the third is the window on the icon not being the
# one that crossed, where the sentence loses its scope (T274); the
# fourth is the icon's figure moved onto the window that crossed,
# reading 0% left (T320). `stoppedelsewhere` is the third one again
# with nothing paying past the quota, where the projection used to
# read "on track" over work that would not run (T288) - and it is
# the row to check a translation against, because the natural
# sentence does not fit in es or fr and the compact rung is what
# carries them. The surface no
# capture can photograph - the shell draws it - so this is how it
# gets reviewed at all. Add --lang to read it in a translation.
--simulate-reset [variant] # a toast card on screen: unexpected (the default early weekly reset)
# | scheduled | credit | session | context | extra | profile |
# profile-failed | extra-bare. The last is the extra-usage card as
# the measured spell reports it - in use, no figure, and so no bar
# at all, since 1 - 0 draws a full one (T277). One table with
# --capture-toast (T198), so a name it does not know prints the
# catalogue and exits 1 rather than showing the default card.
```
## The captures (off-screen, deterministic)
Prefer these over `scripts\Capture-Window.ps1`: that one copies the pixels **on screen** inside the
window's rectangle, so any app that steals focus or sits on top lands in the file. The exception is a
popup — its own top-level window, which `RenderTargetBitmap` over a page's content cannot see.
**Capturing a popup: pass `-Expect <surface>`** (T217). A popup is its own top-level window, so copying
the right window proves nothing about the popup being in it — three captures of the method note came
back green, named the right window, and showed no note. The preview prints a `preview-surface:` line
saying what it drew and where; `-Expect` makes the script demand that line and prove the rectangle is
inside the copy, writing nothing if either fails. The surface for `--stats method` is `method-note`.
Preview popups are held open by `PageWindow` for every popup, not per call site.
```
--capture-settings <out.png> [page] [card=<x:Name> | scroll=<dip>] [profile=<n>] [--sample] [--reveal]
# page System REQUIRES --sample and is refused without it (T205):
# it renders this machine's real login, and a capture is a file
# that gets published. Interactive --settings System still shows
# the real account - only the path that writes a file is refused.
# card= is the FRAME, and the one to reach for (T375): it scrolls
# that element's top into view and prints
# capture-frame: <name> scroll=<n> element=<h>dip
# viewport=<v>dip WHOLE | PARTIAL <pct> of it
# so a card taller than the window says so on the FIRST capture
# instead of the fourth. An unknown name writes no file and
# exits 1, listing the page's own named frames (template parts
# excluded by TemplatedParent, not by a list of ids). Any x:Name
# on the page works; the list shows those over 32dip.
# scroll=<dip> is the old form and goes stale the moment the page
# above the target gains a line. Passing BOTH is refused: two
# answers about one frame, and picking either silently is how a
# shot of the wrong region gets published.
--capture-stats [outBase] [variant] [modifiers] [profile=<n>[,<n>]] [--sample]
# REFUSES, writing no file and exiting 1, when the report has not
# finished (T298): the settle is a fixed 2.5s and the pace is
# computed on a task, so on a slower machine this photographed
# "Computing your consumption pace..." and announced four PNGs -
# twice in five runs. `refresh` is exempt, since capturing
# mid-recompute is exactly what that modifier is for (T118).
# OMITS the Sessions PNG, names it on stderr and exits 1, when the
# transcript scan has not landed either (T343). One pane of four,
# so the three that rendered are still written - read the `wrote`
# line, which lists what exists rather than a fixed four. NOT
# exempt under `refresh`: that asks for a mid-recompute pace
# page, not an unfinished scan.
# all four tabs -> <outBase>-5h.png / -7d.png / -throughput.png /
# -sessions.png.
# profile=1,0 walks the picker one settle apart (the T164 round
# trip). Refuses a variant it cannot show rather than capturing
# the default one (T186).
# Whose name sits above the chart (T197): --sample fills the picker
# from AccountFixture, the default variant from this machine, and
# a synthetic variant on its own gets no picker at all. So any
# PUBLISHED capture of a fixture week needs --sample, and profile=
# is refused where there is no picker to walk.
--capture-tooltip <out.png> [variant] # the same tooltip text drawn into the shell's card, transparent
# background, same variants as --tooltip (T214). A RENDERING, not
# a photograph: card only, no taskbar and no desktop. This is what
# docs\tooltip.png comes from - it used to be a hand-taken shot of
# one machine's notification area, which is why it went a release
# out of date without anything noticing.
--capture-toast <variant> <out.png> # one toast card + shadow + confetti, transparent background.
# BOTH arguments required (T198): a lone path used to be read as
# the variant and land the default card in the working directory.
# Variants are one table, src\Cli\ToastPreviews.cs, which
# --simulate-reset reads too; an unknown name prints the catalogue
# and exits 1.
--check-toasts # every card, asked TWO things in the language this process is in:
# whether it FITS its own frame (T257), and whether it drew a
# quota meter exactly where its row's reading carries a quantity
# (T291). Exit 1 naming the card and either the rectangle or the
# bar. The second question needed its own asking: `Overflow` walks
# TextBlocks and the meter is a Border, so a card drawing one for
# nothing fits perfectly well - and that picture is a full
# allowance behind a sentence saying the quota is spent.
# `--lang <code> --check-toasts`, five times, is the
# whole sweep - the language is fixed per process - and check.yml
# runs exactly that on every push. Reads a layout, not a picture,
# so there is no settle wait and no window is activated.
--render [dir] # tray-icon PNGs at 16/20/32 px, plus TWO contact sheets, each
# magnified 8x with the real pixels preserved and drawn on a
# light and a dark strip: mark_sheet.png for the profile accent
# band (T147), billing_sheet.png for stopped-against-paying at
# each size (T265, the question IconRenderer names as the one
# worth asking). Two files, not two rows of one: a single image
# asserting two claims cannot say which of them failed.
--makeicon [ClaudeTray.ico] # regenerate the multi-resolution app icon
--social [build\social-preview.png] # the 1280x640 GitHub repository card
```
Every one of these creates the directory it is given (T187), so a fresh output folder is fine.
**Where an omitted output path goes** (T198). A default output belongs somewhere git ignores, never in the
working directory — which for anyone running these is the repository root, where `run-commit.cmd` stages
everything. So `--capture-stats`, `--render` and `--context-report` default under `docs\_preview\`;
`--capture-toast` has no default at all. The two exceptions are deliberate: `--makeicon` and `--social`
each default to their own **tracked** artifact, so the default means "regenerate the committed file".
## The headless read-outs
No window, no screen — the arithmetic and the readings as text.
```
--selftest [--quick] # the whole self-check, exit 1 on failure. Two kinds of claim: an
# invariant over synthetic inputs, and a document held against
# the thing it documents — this catalogue included, so a flag
# added here and nowhere else goes red (T243). The second kind
# reads repository files, so an installed copy skips it by name.
# --quick skips the tail sections, which wait on real sweeps, and
# the linking-script run (T372), which is three powershell.exe
# launches against a throwaway pair of config dirs.
--profiles [dir …] [--check] # every profile discovery finds, its auth, its config-dir action and
# its icon accent; extra dirs are treated as registered ones, and
# --check also asks `claude auth status --json` per profile. Also
# prints each profile's last turn, where auto-follow points, what a
# pick in the menu reaches — tray only, or Windows too (T171) — and
# which profile the environment selects beside the one the icon
# draws, marked when the two DIFFER (T172), and whether the tray
# claims to own the variable and what it restores to (T173) —
# that claim against a differing registry is a write that was
# accepted and never landed.
--link-profiles <a> <b> [out=<path>] # the script that makes two profiles ONE SETUP (T367). <a> keeps its
# real files and receives every merge, <b> becomes links into it;
# each side is an index into the --profiles list or a config dir
# outright. Prints the per-entry plan — merged, adopted whole,
# withheld, never — then the PowerShell itself; out= writes it to a
# file instead. The app never runs it and has no write path into
# ~\.claude (§I.4): the script prints its plan on a bare run and
# needs -Apply to act, never elevates, never deletes, and refuses
# before the first move when it needs a file symlink on a machine
# with no Developer Mode. Exits 1 on a bad index or the same dir
# twice.
--probe [--live | --recorded] [--all] # the rate-limit headers VERBATIM: the recorded capture log first,
# then one live call — which is itself recorded (T212), against
# the monitored profile and filed under its key. --live skips
# *reading* the log, not writing to it; --recorded is its
# opposite half (T226) and skips the CALL — reviewing what was
# captured no longer spends a request against the very account
# being measured, which is the account whose requests are worth
# something. Not the default, deliberately: a stale log read as
# if it were current is how a wrong reading gets quoted. The two
# together refuse and exit 1. --all reads every profile's log and
# names which single profile the live call refreshes, with each
# column's last recorded time, so a spread whose profiles are not
# equally fresh no longer reads as one that is. A live reading
# that is kept says the summaries above predate it and names
# --recorded, which re-reads them for free. The instrument for
# T181's other half — what 100% of
# the overage window amounts to, which only an account in overage
# can answer. Each profile's log opens with its VOCABULARY (T210):
# every value each categorical header has taken, counted and dated,
# so "has a second value arrived?" — the question §XVIII.9 and
# §XVIII.10 are blocked on — is read rather than eyeballed across
# 500 readings. A name not sent on every reading says so, and
# --all ends with the CROSS-PROFILE spread (T211): which names
# every account is sent and which only some — the comparison that
# measured `unified-fallback` as conditional and absent, not
# universal. Every printed name is marked READ or UNREAD (T278) —
# whether it reaches a field in this app, taken from ApiClient's
# parser enumerating itself, never from a table beside it. Each
# profile opens with a READERSHIP count, and both directions are
# named: a recorded name nothing reads is permitted but loud, and
# a name the parser reads that a profile is never sent is the
# opposite fact. Quota metadata only: no message content, no token.
--insights # the 24h usage breakdown: requests, sessions, per-model share
--tail # every assistant turn as it lands, with what the sweep cost
--live [seconds] [--root <dir>] [--raw]
# the rolling tok/s with its per-project sparklines, one line a
# second so the metric can be watched against real work;
# seconds bounds the run (default 90), and --raw prints the
# unsmoothed box filter beside it, which is what the attack-only
# smoothing is judged against
--sessions <session-id-or-prefix> # ONE conversation as its CALL TREE instead of the list: task →
# workflow → agent, each node's own cost beside its subtree's,
# which is the reading that says whether the coordinator or the
# fleet was expensive. A task is cut at each person-ask; a
# typed prompt is named by its length, never its text
--sessions [--all|--refresh] [--project <slug-or-name>] [--root <dir>]
# one row per CONVERSATION — the unit no other reader produces:
# last turn, duration, calls, billed tokens, cache reads, how
# many transcripts the fan-out wrote, models. Newest 20 unless
# --all; --refresh re-reads every transcript past the per-file
# cache, which is the only way a wrong cached row is falsifiable
--activity [--numbers|--refresh|--measured|--fold] [--root <dir>]
# the weekly activity shape behind the projection as a 24x7 grid;
# --measured is the same week out of the folded hourly store and
# --fold folds every complete day into it now
--context [slug|name] # what a session costs before you type: every project, or one
# source by source (eager/lazy, bytes, ~tokens)
--context --all # every project in full detail
--context --calibrate # the estimate against transcript-measured session zero, and the fit
--context --check # the rule engine's findings, grouped by severity
--context --usage # which skills/agents were actually invoked (90d)
--context --skills # expand the folded skill/agent index instead of one summary row
--context --prompt [project] # the cleanup prompt the window copies
--context --no-cache # force a cold scan, skipping the %LocalAppData% cache
--context --root <dir> # scan a fixture tree instead of ~/.claude
--context --sample # build and scan the bundled fixture, where every rule fires
--context-report [out.md] # the whole picture as one markdown file
```
## Running a tray beside the resident one
```
--second-tray # skip the single-instance mutex and run a SECOND tray, so a check
# can drive the build you just compiled without you quitting the
# app first (T237). Explicit only — never inferred, and the mutex
# is still taken when free, so ordinary launches keep exiting.
# The tooltip is tagged `[check] …` because notification-area
# icons belong to the shell, not to us: pid cannot tell two
# ClaudeTray icons apart, so the tooltip has to. Match it
# UNANCHORED — the shell's accessible name is the registration
# name followed by the live tooltip, so the tag is mid-string.
# `Check-Interaction.ps1 -Case Menu` uses this on its own when it
# finds a tray already running.
# It POLLS AND DRAWS but WRITES NOTHING (T239): no store append,
# no cache, no settings save, and CLAUDE_CONFIG_DIR is sampled so
# the reconcile lands on a copy. Reads are untouched, or it would
# render numbers no user's tray renders. `--selftest` asserts the
# whole of %LocalAppData%\ClaudeTray is unchanged after driving
# every writer — a new store must consult `ProfileStore.Observing`.
```
## Three flags that go with anything
```
--lang <code> # render this run in en | pt-BR | pt-PT | fr | es | auto, whatever
# the saved preference is. Both tokens are stripped before
# anything else parses, so it can go anywhere. Published
# screenshots are English; use it to check a layout in the
# longest translation. A code this build does not ship prints
# the catalogue and exits 1, and so does the flag with no code
# after it (T260) - it is the flag the whole i18n loop rests on,
# so falling through to the machine's language would write a
# capture of the wrong one under the name the caller gave.
# Matched exactly: `EN` and `pt_BR` are refusals, not pt-BR.
--sample # feed a surface its fixture instead of this machine's data. Any
# published screenshot of a page that shows an account, a repo
# name or a project path must use it.
--sample-env <mode> # answer as if CLAUDE_CONFIG_DIR said this, and WRITE NOTHING (T231).
# agrees | it names the profile the icon follows — the ordinary
# state, which this machine is always in anyway
# other | it names a REGISTERED profile that is not the icon's,
# which is the only way to see T172's "set in Windows"
# outside | it names a folder no profile covers — T172's own line
# unset | no variable at all, so the default ~/.claude applies
# Stripped like --lang, applied before anything reads the variable.
# The mode picks from the profiles really registered here, so the
# fixture cannot produce a menu state that could not occur; an
# unknown one prints the catalogue and exits 1. Sampling is one-way
# for the process — that is what makes "writes nothing" true — and
# `--profiles` says out loud that it is answering off a fixture.
# `Check-Interaction.ps1 -Case Menu` SWEEPS `other` and `outside`
# on its own, one tray each, so T172's mark and line are asserted
# with no flag typed (T238) — a fixture nothing routine reaches is
# coverage nobody has. `-SampleEnv <mode>` pins the run to one
# instead; both it and the sweep are refused with -UseRunning.
```
## Adding a flag
Route it in `src/Tray/Program.cs` (`Main` is the arg dispatch and nothing else), put the work in a
`src/Cli/*Cli.cs`, and **write it down here** — plus a row in AGENTS.md's file map if it comes with a
new file. A flag nobody can find is a flag that gets written twice.
Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.
No comments yet. Be the first to comment!
