An agent sandbox bind-mounts /dev/null over protected paths such as
.bashrc and files inside .claude/, so make patch refused a tree that
was clean outside the sandbox. Untracked character devices cannot be
committed, so the check skips them and lists only what blocks the
release. Untracked files are listed one by one rather than with
--directory, since the collapsed .claude/ entry is not a device.
Work already on the default branch cannot go back to committed without
reopening shipped code. The reviewer files each remaining issue as a new
ticket, related issues together, comments why the ticket is closing, and
moves it to done.
`:mv --status ` offered the five statuses on the board and `:mv -s `
offered nothing: argument_at read the typed word with strip_prefix("--"),
so a word spelled with one hyphen was not a flag to it and fell through to
the positional arm, which counted a position that was not there.
long_name translates the letter through the command that holds the flag, so
the tables stay keyed on the long name and nothing here keeps a second list
of every option's letter. One letter, never a cluster: `mv -fs done` would
make the last letter the honest answer, but --title, --acceptance and a
comment body all take a leading hyphen on purpose, and a criterion typed as
`-fixes the offset` must not be answered with statuses.
The fish half had the same hole in its dynamic rules alone, which were
written as `-l <flag>`: `append -g <TAB>` offered nothing where
`append --tags <TAB>` offered the board's tags. fish_option reads the letter
off the definition per command, so an option given one tomorrow is offered
under it the day it is given. Clap's own generated rules already carried
both spellings, which is why a shell answered `mv -s ` and the board did
not.
a_short_flag_answers_what_its_long_spelling_answers walks every option with
both spellings on every command and nested command and pins that the two
answer the same, so this cannot come back one flag at a time.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The reader knows the grammar approximately: that mv takes a status, not
whether it is a positional or --status, and that set has a flag per section,
not whether the criteria one is --acceptance or --criteria. In a shell fish
answers that as the line is typed. On the board there was nothing, and
getting it wrong cost a rejected line and a guess.
usage_of walks the typed verb against the clap definition and renders its
usage, so a flag added to a command is described here the day it is added.
A table of synopses kept beside it would be a second grammar, and it would
go stale silently: nothing but a reader typing that verb would ever see the
two disagree. The Usage: label is dropped, being the name of a block in a
help page of several where this is one line beside the line it describes.
It takes the keys' slot. esc and enter are two keys a reader learns once,
against a synopsis that is different for every verb; a refusal still wins
over both while it is up, being what was just said about the line that is
still on the row. refusal is the half that says what is wrong where this
says what would have been right.
A verb that resolves to nothing draws nothing and complains about nothing,
since the line is half typed most of the time the row is looked at. A
group's child is walked when the second word names one, so campaign set
describes set; while that word is half typed the group's own usage stands.
Grey, and cut with an ellipsis rather than wrapped: the footer is one row
and a second would reflow the columns every time a verb resolved. There is
always at least the keys' width to cut into, because that width comes off
the row before the line is windowed, so the usage never takes a cell from
what is being typed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
campaign ls printed every campaign and offered no way to narrow it, and the
row was assembled by hand, so a long title pushed the standing out of line
with the row above it and the created date the file already carries was not
shown at all.
--state takes open, done or empty and matches against Standing::state(),
so the derived value keeps one definition: a campaign reads [open] in the
listing if and only if --state open lists it. --state rather than --status,
which names one of the five ticket statuses everywhere else and would
invite --state backlog here. empty is the third value because an empty
campaign derives None deliberately, and a filter that could not reach those
would leave what has been filed and forgotten answerable only by reading
the whole listing. A filter that matches nothing names the value back,
since a silent listing reads as the other empty case.
The filter reaches --ids as well, the way ls --ids keeps its own, and is
the only thing that makes that path count members: an unfiltered completion
is as cheap as it was.
The columns are measured on the raw values and padded before anything is
painted, the rule metadata_cells follows. The codename is measured with the
title rather than given a column of its own: a campaign filed without one
spends nothing on it, so a column for it would be one most rows pad into
and some leave empty, and what a reader lines up is where a name ends and
its progress begins.
A flag's possible values now come off the command that holds the flag, so
campaign ls --state reaches ls's own arguments rather than campaign's. The
prompt offered nothing there while fish offered the three values, which
made a liar of the rule that a ValueEnum added anywhere is typeable on the
board the day it is added.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
WORKLANE_ACTOR is set once in the environment an agent runs in, which is
right for an agent and wrong for the board: a person who exports it where
their own shell picks it up, the obvious way to configure it for the agents
on that machine, then moved a ticket with ':' and the board recorded the
move as the agent's.
Nothing but the reader can type that line, so the prompt asks for
--actor-from-git's precedence without the flag. An explicit --actor still
wins, which is how the reader puts somebody else's name on their own
keystrokes, and a machine with no git user configured reads unknown rather
than falling back to the variable, for the reason the flag does.
The attribution is a parameter of dispatch rather than a field the prompt
writes into what clap parsed: rewriting the arguments would need an arm per
entry-writing command, and the command added after it would be attributed
to the agent until somebody noticed. no_handler_decides_its_own_attribution
reads src for a handler naming an Attribution of its own, which the
compiler cannot catch because it compiles.
The same command typed in a shell is unchanged, which is what keeps a board
several agents write to saying which of them wrote what.
Also closes a code fence in the skill that e26e589 left open, duplicating
the mv --actor example under it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
With the ticket suggested, the rest of the line was still typed blind. The
prompt now completes everything fish does: the verb and the verbs of a
group, the five statuses, the tags the board carries, the campaigns a
codename resolves against, and relate --clear's three kinds.
`ticket_argument` becomes `argument_at`, which answers what a position
names rather than whether it is a ticket. The tables say Ticket, Campaign
or Tag; clap says the rest, so the verbs, the statuses, `set --clear`'s
fields and the `completions` shells are never written down twice. Only
`relate --clear` needed a table of its own: the flag takes a kind and then
a target, so possible values on it would refuse every target.
The row draws the longest opening every candidate agrees on, which is
exactly what the next tab takes. One rule for one candidate and for many: a
ticket position has only the selected card, so the rest of its id is drawn
as before; the five statuses agree on nothing, so nothing is drawn and tab
steps through them one press at a time.
The walk holds the line as it stood before the first press, not the list
and not the candidate the last press chose, so every press re-derives the
candidates and stepping cannot narrow them to what it just typed. Every
edit ends it, beside `refused` and for the same reason.
193f9be is not inherited: `tags set` is offered only the tags the
vocabulary declares, and `tags add` and both `--tags` that take the whole
list are offered nothing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The reader has the ticket selected, which is the whole reason to type a
command on the board rather than in a shell, so where the line is naming a
ticket the selected card's short id is drawn after the cursor in grey and
tab takes it into the line.
Where a ticket goes is completions.rs's answer, through a new
`ticket_argument` over the three tables the fish half is generated from. A
copy of the positions on the board would go wrong the next time a command
grew a ticket argument, and silently, since the tests guarding the first
copy would not read the second.
TICKET_FLAGS holds (command, flag) pairs now, the shape TAG_FLAGS already
had, and the `chown --clear` rule the fish block wrote out by hand is
generated from it. Only chown's --clear names a ticket: relate's takes a
kind and set's a field, and the flag alone cannot tell the three apart.
The suggestion is derived on every call rather than held on the Prompt, so
what the row draws and what tab takes are one answer, and it is offered
only with the cursor at the end of the line, while a card is selected, and
while the id still starts with what is typed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
check_blockers reported a blocker the board does not hold with eprintln!,
the one print site fa5662a left behind: a helper had no sink to reach, and
what it wrote was stderr rather than stdout. Harmless while main was the
only caller, and since 53ffb94 the command prompt runs the same handlers,
where stderr is the alternate screen with the columns on it. ratatui does
not repaint a cell it believes is unchanged, so the fragment stayed until
something else redrew them.
Sink carries two writers instead of one, because the CLI's two streams are
two: Sink::std hands the answer to stdout and an aside to stderr, so a
shell reads the same line on the same stream, and the integration test now
asserts that stderr whole rather than searching it. Sink::buffer holds its
warnings instead, and Board::run reads them back with Sink::warnings and
puts them above the answer on both outcomes -- a mv that warns about a
dangling blocker and is then refused by a real one wrote the aside first,
and it says why the refusal names one blocker and the ticket records two.
check_blockers and check_dependents both take the sink now, so their
refusals take the colour from it as well, rather than asking the process
about a stdout the board is not writing to.
no_handler_or_helper_reaches_a_print_macro walks src for the next print
site, each file cut at its own test module, since the rule is about every
future site rather than the one that broke.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
check_blockers refused every move while a blocked by target was still open, so a
dependent sat in backlog until its blocker landed even when the blocker was in
reviewing and its code was there to build on. The way out was -f, which also waives
the backward and skipping guard, on every move for as long as the pair existed.
The rule is now the columns: a ticket may stand level with what it waits on but never
ahead of it, so a blocker in reviewing lets the dependent be committed and reviewed
and gates only done. util::blocks_move is the one place it is decided, taking the
status moved into and the blocker's own, and mv passes the target through. A finished
blocker still holds up nothing, answered through is_open, and a target of discarded
is answered by name rather than by its place in Status::ALL, where it sits last.
The refusal now names the status moved into and only the blockers still behind it: a
dependent told to wait for work already ahead of it has nothing to act on. The board
asks the same function about the column one stage ahead, so the blocked mark means the
card cannot advance rather than that a dependency was ever recorded, which also takes
it off cards in the two terminal columns.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>