install-skill now writes every document the binary carries, with an optional
name to narrow it. Requiring a name per document would mean a skill added
later never reaches anyone who already knows the invocation, which is the
drift the command exists to prevent. --print is the exception and requires
the name: a diff is against one document.
The review skill is its own document rather than a section of SKILL.md,
because the frontmatter description is the whole of what decides loading and
one description cannot name both "anything touching a ticket" and "only when
a review is asked for".
The reviewer never ticks a criterion and never moves the ticket. A tick is
the implementer's claim and unticking is the review disagreeing; one mark for
both loses the disagreement. mv belongs to whoever does the work.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
There was no way to group the tickets that have to be finished together. A
goal such as a v1 release was either one ticket that could not describe its
parts, or a set of tickets with the goal restated in every description and
nothing tying them together.
A campaign is a file of its own under `.worklane/campaigns/`, with a title, a
description and metadata, and it sits in no status directory. Its state is
derived instead: open until every member is done. Deriving it means there is
no such thing as a done campaign holding an open ticket, and no second
forward-only ruleset beside `mv`'s. An empty campaign is neither open nor
done, since calling it done would make a container nothing has been filed
against the most complete thing on the board.
Membership is one field on the *member*, Campaign in the ticket Metadata,
written only by `assign`. The member list is deliberately not on the campaign:
that would put membership in two places and make one file that every
assignment edits, contended the moment two agents share a board. So `add`
gains no `--campaign` and neither does `set`.
`add` does not write the field either. Every ticket written before campaigns
existed has none, so the insertion path has to exist regardless, and leaving
it out of `add` means one path that is always exercised rather than two. That
is the Relations argument from 4179958, applied to a field instead of a
section. `set_or_insert_metadata_field` is the upsert and `METADATA_ORDER` is
its anchor list, the way `SECTION_ORDER` is `ensure_section`'s, so Campaign
lands beside Status rather than at the end and Status stays last. A file with
no Metadata heading is an error rather than an insertion, per 641e3d5.
The three tiers came out of `resolve_ticket` into `resolve_in`, shared by both
kinds. A campaign is addressed exactly the way a ticket is, and a second
resolver would be a second copy of the precedence rules to keep in step. Two
things fell out of that. `Found.dir` is the status for a ticket and empty for
a campaign, which is what drops the bracket from the ambiguity listing for
one. And `resolve_ticket` can now explain a search that names a campaign, so
`mv`, `show`, `chown` and everything else that takes a ticket give the same
reason instead of a bare "no ticket found" — one place rather than a check per
command. Only on the zero-match path, so an ambiguous ticket error is never
replaced by a complaint about a directory the caller never mentioned.
The stored value is the full uuid, per the same rule the Relations section
follows, and the title is resolved for display. `reference_cell` is now shared
between the two, so a dangling campaign and a dangling relation target are
both `_missing_` rather than each inventing a spelling. `Resolved` carries
what `content_trim` cannot render from the file alone: a struct rather than
two more parameters, so the next such field is one entry instead of an edit at
every call site.
`campaign show` lists done members, unlike `ls`, which leaves the done column
out of its listing. What is finished against what is left is the whole
question a campaign asks. `ls --campaign` is the other way round and keeps
every rule `ls` already has.
Deliberately out of scope. The derived progress figure is 1cb999f, now
recorded as blocked by this, since nothing here can compute it yet.
Unassigning needs a field cleared back to unset, which is 74904dc. And the
board gets neither a campaign tab nor a card marker: the detail view resolves
the field because it renders the body, and that is all. A marker would want a
read per reload and the four corners are already spent, which is the b8ce18c
reasoning; filed as 61879a4.
Twenty-four integration tests and eleven unit tests, covering assign,
reassign, the no-op, a dangling campaign, the ls filter, `mv` refusing a
campaign, insertion placement, and a handwritten section and unknown metadata
field surviving an assignment verbatim.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`relate` is the only writer of the section, the way `chown` is the only writer of
Owner: a relation names a target that has to be resolved to a uuid first, so `set`
and `append` are deliberately kept out of it. One flag per kind, so an unknown kind
is a usage error rather than a line refused on the way back out.
One direction is stored and the inverse is derived by a scan. `blocked by` goes on
the gated ticket, whose move `mv` now refuses while the target is not done, naming
every open blocker; `-f` bypasses it beside the backward and skipping moves. A
target that is not on the board warns and allows the move.
The rendered block is built outside `content_trim` and passed in, since half of it
comes from other tickets. The section is created on demand rather than by `add`, so
a ticket written before it existed gains it without being regenerated.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The Owner field had no writer after `add`. It was folded into `set` in da03690
and comes out here as its own command, so no field is writable from two places,
the same reasoning that keeps status in `mv` alone.
`worklane chown alice 641e3d5`, user first, after Unix `chown`. That reverses
the ticket-first order `mv`, `comment` and `show` use, deliberately: the name
borrows the Unix association, so the argument order borrows it too.
The swap is left to resolution rather than checked for. Given `chown 641e3d5
alice` the second argument has to resolve as a ticket, and a bare username does
not, so the call fails on its own; a check on the first argument would only fire
where resolution already refuses and would stay silent on the one case it cannot
catch, a ticket whose title is the username. `cmd_chown` is therefore
`set_metadata_field` plus `save_ticket`, with the whole-file no-op comparison
`set` uses and the same `report()` line, so an owner set to the value it already
holds is reported unchanged and nothing is written.
Nothing new writes a file: the update goes through `write_then_replace`, scoped
to the lines under `## Metadata`, so a description quoting `**Owner**:` is left
as prose and a failure leaves the ticket whole. Tested for the change, the
no-op, all three resolution tiers, a swapped call, an empty user, a ticket with
no Owner field, and that the tick, the handwritten section and the filename all
survive.
The skill and the opencode tool carry it too, the tool with the argument order
in its description, since that is the one place a model can get it wrong
without failing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Done only grows. It already holds more tickets than the other three columns
together, so a quarter of the width was going to the column least often read
while the ones being worked got squeezed for it.
Not a new opinion: `ls` already counts done in its summary and leaves it out of
the listing below. The board now does the same, and the footer carries the
count while the column is hidden, which is the board's version of that summary
line. It reads `d show done (17)` while hidden and `d hide done` while shown,
so the count only appears where a column heading is not already carrying it.
The modal footer names the key too, since it works with a ticket open.
Hiding is a change to what is drawn, never to what is read. `reload` still
scans the done directory and `digest` still watches it, so the count is right
and revealing the column needs no rescan.
`visible()` returns a prefix of the columns rather than a filter, which is what
keeps `column` and `offsets` valid as indices into either. It is derived from
where DONE actually sits rather than assuming the last slot, and a test pins
done as last: a status inserted after it would otherwise hide two columns in
silence. DONE is now a constant, since `ls` and the board are two readers of
the same rule.
Hiding the column under the selection moves the selection to the last drawn
column and closes an open detail view. Refreshing the modal onto whichever
ticket the selection landed on was the alternative, and it swaps what someone
is reading without saying so, which is how a person acts on the wrong ticket.
Showing done again is a no-op for the selection, since a column is appended
rather than inserted.
Three existing tests asserted a column heading with `screen.contains("done
(0)")` over the whole render, and the new footer holds that string whether the
column is drawn or not, so all three passed for the wrong reason the moment
done was hidden. They now assert against the top border row through a
`headings()` helper, with `footer()` beside it.
The choice is not remembered between runs: worklane has no config file, and one
flag is not a reason to start one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`cargo clippy --all-targets` was reporting 96 warnings, 14 of them in the
binary. It is now clean from a cold cache, and the 211 tests still pass.
`clippy --fix` took the mechanical ones: needless borrows, `map_or` that is
`is_some_and`, `push_str` of a single character, manual `split_once`, and
`&PathBuf` parameters in the test fixtures. It also moved the `handlers.rs`
test module to the end of the file, which is what `items_after_test_module`
was pointing at.
Three by hand. `cmd_add` took nine parameters threaded from `main`, and now
takes `AddArgs`, which is how `set`, `append` and `check` already pass theirs;
the handler destructures it so the body reads as before, and `add --help` is
unchanged. The collapsed `if` in `content_trim` needed reindenting, since
`--fix` left its body at the old depth. And the two `split_once` rewrites read
as `.split_once(..).map(|(_, value)| value)` rather than clippy's `.map(|x|
x.1)`.
The two eight-argument `create_ticket` test fixtures keep an
`allow(too_many_arguments)` carrying the reason: the lint is right that
twenty-five positional call sites read as a wall of strings, but replacing them
with a named-field struct is a refactor and not a warning sweep. a34a316 tracks
it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
worklane's agent instructions only existed for someone with the repo checked
out. Anyone who installed the binary with cargo install got no document at
all, because cargo install places binaries and nothing else.
The skill now lives at assets/SKILL.md and is embedded with include_str, so an
installed copy can never be a different version from the binary that describes
it. `worklane install-skill` writes it, `--print` sends it to stdout so a
modified copy can be diffed, and a copy that was edited is reported and left
alone unless --force: it may have been changed on purpose, and silently
discarding that is worse than stopping.
One path serves both agents, ~/.claude/skills/worklane/SKILL.md. That was
settled empirically rather than from documentation: opencode 1.15.10 loads
~/.claude/skills as an external skill source alongside its own
~/.config/opencode/skill, and with HOME pointed at a temp directory holding
only what install-skill wrote, `opencode debug skill` lists the document under
the name worklane. Writing opencode's own path as well would be two copies of
one document, free to drift, for no gain.
The document is new rather than the one under .opencode, which is the user's
and not this repo's to ship. The ticket-workflow prose moved out of AGENTS.md
into it, so how to use worklane has one writer, and AGENTS.md points there.
The mechanism is in install.rs rather than skill.rs, and install() takes its
root as an argument: install-tool is coming, it needs a different root and the
same compare-and-refuse logic, and passing the root is what lets the tests
install into a temp directory instead of the home of whoever runs the suite.
Three tests guard the document itself, which is the part that goes stale
quietly: the frontmatter has to parse, every `worklane <x>` it shows has to be
a real subcommand, and every subcommand outside NOT_FOR_AGENTS has to appear in
it, so a new command cannot ship invisible to an agent.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
worklane was one-shot commands only, so reading the board meant running ls again
and again and nothing ever showed the four statuses side by side, which is the
one view a kanban flow is for.
Four columns, a card per ticket carrying the priority mark, short id, title,
owner and due. Arrows or hjkl to move, enter for the detail view, q to quit.
Read only: moving a card by keypress is a follow-up, kept out until the watch
and render loop were solid.
Modelled on frig throughout, same crates and same shape: event poll on a 250ms
timeout so the tick runs when nobody is typing, a panic hook that leaves the
alternate screen before the backtrace, and the terminal restored on the error
path as well as on quit. A bare worklane opens it, as a bare frig does, so
Cli::command is now an Option.
The watch looks at files, not directories. A directory's mtime only moves on
create and rename, so a ticket edited in place would be missed; the digest
hashes every ticket and comment file's path, length and mtime. Comment files
are in it because a comment added elsewhere has to reach the detail view. It
guards the rescan rather than the draw, since ratatui diffs its buffer but
reading the tree every frame would not be free. notify is deliberately not a
dependency, and the reason is on the ticket.
Two orderings are deliberate. The worklane directory is looked for before the
terminal, so running outside one gives the usual message rather than a complaint
about the tty, and both errors stay testable without a pty. And a missing tty is
refused by name, since raw mode on a pipe fails with a bare "No such device or
address" that explains nothing.
The palette is now a Color enum rather than a string, so the escape the CLI
writes and the ratatui colour the board draws come from one definition and a
colour missing from either is a compile error. That retires the test which
existed only to catch a name with no entry, the failure that once left magenta
undefined.
Cards come from collect_ticket_items, so there is no second parser;
TicketItem gained the path the detail view opens. content_trim and
format_comments_with take the colour decision, because ratatui draws spans and
would print an escape as literal bytes.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Criteria are stored as - [ ] lines and nothing could tick one, so marking
progress meant replacing the whole list through set, losing the state of every
other item.
check and uncheck are separate commands rather than one verb with a flag,
sharing CheckArgs and cmd_check. The criterion is an index first and a substring
second, so a criterion whose text is literally "2" is reachable only by
substring. A number out of range is reported as a bad index rather than as a
substring that missed.
Only the character between the brackets is spliced, so the text keeps whatever
indentation and spacing it had. Criteria are counted inside the Acceptance
Criteria section only, so a checkbox written into a description is not a
criterion and a handwritten line in the section takes no number.
With the criterion omitted, the criteria are listed and the number asked for
when stdin is a terminal. Where it is not, the same list comes back as an error
instead. Prompting regardless would leave the opencode wrapper blocked on a
stdin that never closes, and that is the choice resolve_ticket already makes for
an ambiguous match, for the same reason: the wrapper keeps stderr and throws on
the code, so a list printed to stdout would be discarded.
Any listing of a subset keeps the real numbers. Renumbering the matches from one
would contradict the advice to use the number and point at the wrong criterion.
show now numbers the criteria as "1. [x] text", so the index is visible where
the ticket is read. That also drops a manual prefix strip, so clippy is one
warning below where it started.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Nothing could change a ticket after it was created. mv wrote status, comment
appended a file, and that was the whole set of mutations, so correcting a typo
or adding a criterion meant hand editing the markdown.
Two commands rather than the one vague "edit" the ticket described. set replaces
a field, append adds to a section, so the verb carries the meaning instead of a
prefix on seven flag names, and append only offers the three sections that hold
multiple lines. Owner is absent from both, it belongs to the chown command filed
alongside this, and status stays mv's. One field, one writer.
Sections are a table, DESCRIPTION, ACCEPTANCE and PLAN, each carrying a heading,
a label, the placeholder add writes when empty, and a kind. Both commands walk
it, so a section added later is one entry rather than edits in two places.
Everything is targeted line surgery and never a regenerated file, which is what
keeps a ticked criterion ticked and a handwritten section intact when an
unrelated field changes. Appending to a numbered section continues from the
count already there rather than renumbering above it, so a handwritten line
survives exactly as written. Appending to an empty section replaces the
placeholder instead of following it.
An empty value is a paragraph break for a section and a usage error for a
scalar, since it looks like a way to clear a field and clearing is not supported
yet. No-op detection compares the whole file before and after each edit, so a
field set to the value it already holds is reported unchanged and nothing is
written.
Two things came out of building it. move_ticket_file's staged write is extracted
as write_then_replace, now the only way any ticket file is written. And the file
is located by uuid rather than by title: titles are not unique, add allows
duplicates on purpose, and the old title scan meant mv <uuid-of-second> moved
the first ticket sharing its title. Reproduced, fixed, and pinned by a test.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The opencode add tool split the description on newlines and then dropped the
empty ones with filter(Boolean), so every multi paragraph description arrived as
a single paragraph, against the tool's own instruction to separate paragraphs
with newlines.
cmd_add needed no change for this. Joining the values with a single newline is
what makes an empty value a paragraph break, so the capability was already
there and only the wrapper was throwing it away. Verified by hand before
touching anything.
The wrapper now keeps interior blanks and trims only the ends, and only trims
trailing whitespace within a line, since leading whitespace is what makes a
nested list or an indented block survive. --acceptance and --plan keep dropping
blanks: each value there is one criterium or step, so an empty line is spacing.
trim_blank_ends in cmd_add makes the second criterion hold whoever calls it, not
just the wrapper: blank values at either end are dropped so a break never lands
against a section heading, and an all-blank description reads as none. Interior
runs are kept verbatim rather than collapsed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
cmd_mv renamed the ticket into its new directory and only then wrote the updated
content, so a write that failed left the ticket in the target directory still
claiming the old status. Reproduced with a read-only ticket file: the move
reported a bare "Permission denied" and left the file in committed/ carrying
"Status: backlog".
move_ticket_file now stages the content in a temp file, renames it over the
source, and only then renames the file into the target directory. Two atomic
steps, so a partial write is discarded and the original is untouched. The other
obvious order, writing to the target and deleting the source afterwards, was
rejected because it has the ticket in two directories at once.
The worst case left is a ticket still in its old directory with a Status one
move ahead, which the next mv corrects since the directory is what decides the
current status. Tested, including the retry.
The temp file is staged under .worklane/ rather than beside the ticket, so a
leftover from a hard kill is invisible to ticket lookup and to the listing
instead of appearing as a stray untitled ticket. IO errors now name the ticket
and the path rather than propagating bare.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
cmd_show printed the resolved title with its rule, then content_trim_colored
printed the "# " heading again out of the file content, so every show output
carried the title twice.
The body now skips the heading, along with the blank line the template puts
under it so nothing is left where it was. cmd_show keeps the surviving copy and
colours it green, which is where the colour guideline puts a title.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Section rules repeated the character for the byte length of the whole heading
line, so every rule ran three past its heading for the "## " prefix. The title
rule in cmd_show used byte length, which over-draws for a non-ASCII title.
Both now count characters of the text as displayed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Priority lived in tags, as p1 to p3, which ls could neither sort nor
filter on, nothing validated, and a typo like P1 passed silently. ls
sorted by title only, so a backlog of any size gave no ordering signal.
Levels are 1 to 5 with 1 most urgent, matching what the tags meant.
Omitting the flag leaves the field _Not set_, like Due and Owner: there is
no implicit middle level, because a priority nobody chose is worse than
none. clap rejects anything outside the range before a ticket is written.
Priority parses to three states rather than two. A stored value outside
the range is Invalid, kept distinct from Unset, so a typo cannot read as
"no priority": it shows as ? and sorts just ahead of the unset tail.
ls sorts on priority and falls back to the old title sort, and --priority
filters, composing with the status filter in both the grouped and the
filtered listing. TicketItem becomes a struct, since a sixth tuple field
was past reading.
Migrated the 31 tickets already carrying p1 to p3 tags: the level moves to
the field and the tag is dropped.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Comment filenames were stamped to the second, so two comments added in
the same second produced the same name and fs::write truncated the first.
Agents post in bursts, so this was routine data loss: five comments added
back to back left one file.
Stamp microseconds instead, and create the file with create_new rather
than fs::write, so a name that already exists is taken as a collision and
the next zero-padded suffix is used instead of overwriting. The suffix
sorts after the bare stem, so read_comments still orders comments the way
they were written.
The collision branch cannot be reached from the CLI, since two processes
will not share a microsecond, so it has a unit test of its own.
Patch bump: existing second-resolution comment files still read and still
sort ahead of new ones written in the same second, so no migration.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Three places sliced strings by byte offset on values that come from a
filename or a file field, neither of which is validated.
The disambiguation list took &short_id[..7], so two tickets whose names
were shorter than that aborted the process while printing the options.
collect_ticket_items guarded on length but still cut at byte 7, so a
multibyte character straddling the cut panicked. format_created_at read
four fixed offsets after a length check that a multibyte value passes.
Add short_id, which counts characters, and use it in both display paths.
Parse the timestamp with chrono instead of slicing, dropping the column
when the value will not parse. The format string it writes and the one it
reads are now the same constant, so they cannot drift.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
cmd_mv spliced the metadata line by byte offset and wrote the field name
as **-Status**, because the format string re-added the "- " prefix that
the preceding slice already carried. parse_ticket_field could never read
Status back, and the next move failed to find "- **Status**" so it
silently skipped the rewrite, leaving the ticket in one status directory
while its metadata claimed another.
Replace the splicing with set_status_field, a line based rewrite that
keeps every other line verbatim, inserts no stray newline, and errors
when the Metadata section has no Status field. It also accepts the
**-Status** spelling so tickets damaged by the old code are repaired as
they move.
The content is now computed before the rename, so a ticket without a
Status field is never half-moved.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>