Files
marketplace/cli/internal/cmd/AGENTS.md
T
naudachu 8b1b11001a feat: drop the kettle plugin; the binary writes its own skills
The plugin and the binary shipped on two release cadences and nothing on an
operator's machine ever checked that the one they installed described the other.
The generated flag block existed precisely so a renamed flag could not ship with
documentation recommending the old one — and then shipped one version behind the
registry it came from, which is the same bug one hop downstream.

So the prose moved into the binary. `internal/scaffold` embeds every document;
`kettle init` and `kettle gen scaffold` write them into a project's own
`.claude/`. The two cannot disagree because there is one artefact.

The namespace survived the move. A project's skills are flat, so the prefix is
spelled into the directory name (`kettle-issue`); a project's *commands* take
their namespace from a subdirectory, so `commands/kettle/init.md` is still
`/kettle:init`. Four of the six command files are thin pointers at a skill, and
that is what kept ~1,600 lines of `/kettle:…` cross-references true without a
rewrite. `init` and `auth` lost `disable-model-invocation: true` — being a
command is that property — and `auth` now restricts `allowed-tools` so a model
cannot reach `kettle auth add` at all.

`gen scaffold` writes files whole rather than splicing a region. The old
refusal protected somebody's hand-written prose around the block; that prose is
embedded now, so there is none to protect, and preserving local edits would
freeze a project's documentation at whatever version first initialized it.
`--check` warns before an upgrade discards one.

The plugin's `agents-sync.sh` — 141 lines of Python behind a filename that said
`.sh` — became `internal/mirror` and `kettle mirror`. Same seven branches, same
refusal to merge two real files that differ, now with a table test per branch
and a check that a repair converges in one pass. `--hook` is the PreToolUse
form and exits 0 on every path including a panic. It is opt-in per project,
which is strictly narrower than the plugin hook that was on for everybody who
installed it.

`kettle init --interactive` walks a person through the login, the token (read
with the echo off, so it lands in no history and no file), the repository, the
`.claude/` tree and the mirror hook. It refuses a stdin that is not a terminal
and names the flags instead: every question it asks has one, and it performs
nothing itself, so an interactive run and a flag run are one code path.

Two rules that used to be prose are now the binary's: init refuses a linked
worktree and names the main checkout, and writing into an existing
`.claude/settings.json` is refused with the snippet printed rather than
reformatting a file the operator commits.

The scaffold version stamp went to its own `.kettle/scaffold.yaml` rather than
into `config.yaml`, because unknown keys there are a hard error and that file
may be committed and read by whatever build each machine has.

golang.org/x/term becomes a direct dependency; it was already in the tree
indirectly, so no module was added.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-08-12 16:17:24 +05:00

16 KiB

AGENTS.md — internal/cmd

The command tree: flags, receipts, exit codes. The only package that may import every layer below it, and the only one that prints.

cmd/kettle is four lines around cmd.Main(os.Args[1:]) — everything a main usually accumulates lives here instead, because a main package cannot be imported and therefore cannot be tested.

Commands are values

Each command is one register(&Command{…}) in an init(), carrying the metadata a human needs — Short, Long, Examples, Args, Groupin the same struct that carries the code. That is what lets a project's skills be generated from this list: a command whose flags changed cannot ship with documentation that says otherwise.

func init() {
	register(&Command{
		Name:  "tree",
		Group: GroupIssue,
		Args:  "[<id>…]",
		Short: "draw the dependency graph of the local store",
		Long:  `…`,
		Examples: []Example{{"kettle tree", "every root (nothing depends on it)"}},
		Setup: func(fs *flag.FlagSet) func([]string) error {
			out := storeFlag(fs)
			depth := fs.Int("depth", 6, "maximum depth")
			return func(args []string) error {  }
		},
	})
}

Setup registers flags and returns the runner, closing over them. Splitting it that way is what lets Command.Flags() walk a command's flags without running anything — which is how the doc generator reads them.

The tree is flat. kettle new, not kettle issue new: an agent pays for every token of every invocation, and the grouping that matters for reading is carried in Group and only shows up in the docs. Four groups, in presentation order: project, issue, sync, api.

file what is in it
command.go Command, the registry, Main, help rendering, SilentError, Fail, and permute
flags.go storeFlag/storeRoot, wasSet, the repeatable stringList
sync.go syncStart/syncStartExisting, commentsSidecarPath — the shared opening of every tracker command
gen.go kettle gen scaffold: rendering scaffold's documents with the registry's flag tables spliced in
init.go auth.go config.go version.go mirror.go group project. version.go also holds Version, the string a release build stamps in with -ldflags -X
interactive.go kettle init --interactive: the only thing here that reads a terminal
settings.go registering kettle mirror --hook in .claude/settings.json, and refusing to reformat one it did not write
new.go check.go ac.go tree.go index.go evict.go group issue — no network in any of them
pull.go push.go remote.go comment.go close.go labels.go evict_sync.go group sync
api.go group api, alone in it: one request to an endpoint nothing here wraps
cli_test.go builds the binary in TestMain, runs it as a subprocess
sync_pull_test.go sync_write_test.go api_test.go the tracker halves, against fake servers
gen_test.go the generator: the whole tree, determinism, and that a local edit does not survive
mirror_test.go the repair command, and the hook form's promise never to fail a Bash call
init_scaffold_test.go the .claude/ tree, the settings refusal, the worktree refusal, the not-a-terminal refusal
scaffold_coverage_test.go the registry ⇄ documents seam, in the one package that can see both

The fourth group is one command and was still worth naming, because a group is a skill written into a project: api is a subject somebody loads on its own — which endpoint, and does it paginate — and folding it into sync would have put "how do I cut a release" behind a skill about the issue round trip.

mirror is the one command in the tree that has nothing to do with issues, and it is here rather than in a second binary for the reason this repository keeps repeating to itself: a binary holds what can be enforced. Repairing an AGENTS.md/CLAUDE.md pair is a seven-branch filesystem decision with one refusal in it, which is a thing a table test can hold down and a shell script cannot. It was 141 lines of Python behind a filename that said .sh, shipped in the plugin's hooks; when the plugin went, the shell had nowhere to live and Go did. The domain is untouched: mirror imports nothing.

Three conventions every command follows

Flags may come after positionals. The standard flag package stops parsing at the first non-flag argument, so kettle ac <id> --check 3 would hand --check to the command as a positional and tick nothing. permute moves flags forward, using the FlagSet to know whether a flag swallows the next argument; -- ends the permutation. Every other CLI an operator uses interleaves the two, and a tool that silently ignores a flag because of where it was typed is worse than one that rejects it.

Exit codes are three. 0 fine, 2 for a usage problem (unknown command, unparseable flags), 1 for an ordinary failure — printed as kettle <name>: <err> by Main, which is why no command prefixes its own errors. SilentError{Code: 1} is for a command that has already said everything it has to say: check and gen --check use it, because findings went to stdout and a second copy on stderr would be noise.

The store is resolved before a socket is opened. syncStart does that in one place: a command that dialled first would report a network problem for a project that was never initialized, and an operator would go looking at the wrong thing. syncStartExisting adds RequireStore for the commands that read the store rather than create it — push, comment, close, sync-evict — because a missing store is a mistake to report, not a directory to conjure.

There is no --login and no --repo on any sync command bar labels. Which login a project runs under is a fact about the project, stated once by kettle init. That the two could disagree is what the Python version needed a PreToolUse hook to police.

api keeps that rule and needs no flag to: a cross-repository address is an address, so repos/other-owner/other-repo/releases is simply a path with nothing to substitute — {owner} and {repo} are filled in only where they are spelled. Another instance is KETTLE_URL/KETTLE_TOKEN, and a full URL pointing at a host that is not this project's is refused by the transport rather than sent with the token attached. It also resolves the store it never reads, exactly as labels does, so "there is no project here" fails the same way for every command that talks to a tracker.

-X DELETE needs --yes. The only gate of its kind in the tree, and it is here because this is the only command that can delete something that is not an issue — a release, a tag, a branch — from an argument. A flag typed on purpose is an operator's decision; everything else about the request goes out as spelled.

--out is the one flag almost every command has, and an explicit one is used exactly as typed: a relative --out stays relative to the working directory, because that is what the operator asked for.

Usage

kettle help                     # the tree, grouped
kettle help push                # one command in full: flags, defaults, examples

kettle init --interactive           # a person at a terminal, walked through it
kettle init --login noodles --repo owner/name
kettle mirror --check               # AGENTS.md ⇄ CLAUDE.md, everywhere below here
kettle new --type task --title "Wire sqlc into the appclick repo layer"
kettle ac wire-sqlc-appclick --check 3
kettle check --strict           # exit 1 on any error; --strict counts warnings too

kettle pull 42                  # the issue and everything blocking it, any state
kettle push --update wire-sqlc-appclick
kettle sync-evict --dry-run

Every command's own Long text is the reference — it is what kettle help <name> prints and what the generator splices into scaffold's documents. Do not restate a flag table here; it would be a third copy of something already in two places, one of them mechanically checked.

--interactive, and the two things it may never become

interactive.go is the only code in this binary that reads a terminal. Two rules hold it in place, and both are about what it must not turn into.

It answers no question the flags cannot answer. Every prompt has a field in initOptions and therefore a flag — --login, --repo, --scaffold-out, --no-scaffold, --mirror-hook. A capability reachable only by answering a prompt would be a capability no script, no CI run and no agent could ever use.

It performs nothing. It fills the struct in and hands it back, so runInit is one code path: the run that follows an interactive session is byte for byte the run somebody else gets from flags.

It refuses a standard input that is not a terminal, and that refusal is load bearing rather than fussy. A model that reached for --interactive would otherwise block forever on a prompt nobody is going to answer; the error names the flags instead. It is also what makes the one genuinely better thing here possible — term.ReadPassword puts a token in no history, no file and no scrollback, which every other route into this binary can only approximate.

--repo is offered with a guess read out of .git/config, parsed rather than shelled out to. git is one more thing that has to be installed and one more process to misread; a guess is worth what it costs, and anything unparseable is no guess at all.

Writing into .claude/settings.json

settings.go registers kettle mirror --hook on PreToolUse(Bash), and is the only place this binary touches a file the operator owns and commits. It has three outcomes and the third is the whole reason it is not ten lines long:

  • no file — written, hook and all;
  • a file already holding a kettle mirror hook — nothing happens;
  • a file holding something elserefused, with the snippet printed to paste, unless --force-settings.

Go's encoding/json cannot preserve key order, so any merge reformats the whole document. An operator who asked for a documentation hook and got an unrelated diff across a file they share with their team has been badly served; a snippet they paste costs them ten seconds. The refusal is reported on stderr and the run still succeeds — everything before it already happened, and reporting a failure would send somebody looking for damage there is none of.

The registered command carries a command -v kettle guard. That line outlives this binary: an operator who uninstalls kettle would otherwise get command not found on every Bash call, from a hook they set up months ago and have long stopped thinking about.

push and pull, the two halves of one rule

The rule is that the store holds what has not left this machine. Both halves are worth reading in full before either file is touched.

push (push.go) deletes <id>.md and every sidecar under that slug — on create and on --update alike, one rule with no exception, because a PATCH is a push and two rules would put back exactly the question this removes ("is my copy the fresh one?"). The deletion is the last thing that happens, and only after all three of:

  1. the call came back without an error and with a 2xx,
  2. the answer carries a plausible number — on --update, the very number that was PATCHed,
  3. the number → slug ledger has been written.

Network down, non-2xx, an answer that does not confirm the write: the file stays and the run stops. Get the ordering wrong and a slug is lost at exactly the moment the local copy stops being the record, which is why the ledger is written before anything is deleted. A never-pushed origin: local issue is never touched by any of it.

Dependencies go first, in topological order, so a blocker has its number before the issue that names it. An --update can take one extra request with it, because Gitea's edit endpoint carries no labels — when the answer's label set and the issue's disagree the whole set goes up in a PUT, and a warning on stderr says which names moved.

pull (pull.go) is how a pushed issue comes back. Three sources answer "what is this issue called here", in this order: the ledger (the only one that knows what is on disk right now, so it wins), the <!-- kettle:id … --> marker in the tracker-side body, then the slugified title. A marker is taken at its word only when the slug is free; a name already in use is a collision, not an identity, and is uniquified.

Two ways to name what to pull, and they are not the same operation: a key is an address and fetches an issue in any state, while a filter is a query and leaves closed issues out. --limit is on the write, not the selection — it counts what lands in the store, which is why a filtered pull can enumerate far more than it keeps and says so. Blockers come down too, recursively to --depth, and are outside the limit: a blocker is followed because a stored issue named it. A pull overwrites the body — it is a fetch, not a merge — with checkbox state the one exception.

The generator

gen.go writes a project's .claude/ tree: scaffold's embedded documents, with this registry's flag tables spliced into the four that declare a region.

It writes every file whole, and that is a deliberate reversal. The old generator owned a region and left every byte outside it alone, because the prose around the block was somebody's hand-written file in this repository. It is not any more — it is embedded — so there is no hand-written half left to protect, and preserving local edits would freeze a project's documentation at whatever version first initialized it. The markers stay in the output so a reader can still see which half came from the registry.

The principle the old refusal protected did not go away; it moved. Nobody's prose is clobbered because nobody's prose is there — it lives in assets/, next to the code it describes, and --check warns before an upgrade replaces a local edit.

The output is deterministic to the byte — no timestamps, no map iteration — so regenerating something unchanged produces no diff. --check is that property made useful: it writes nothing and exits 1 when anything on disk differs, which is what a pre-commit hook or a CI step calls, and it wins over --dry-run.

One document per group, so adding a group here means adding one under internal/scaffold/assets; name a group only when it is a subject somebody would load on its own. TestEveryGroupHasSomewhereToBeWritten is what makes that a two-step change rather than a silent one-step mistake — the generator walks the documents, not the registry, so a group with no document would have its commands written nowhere and nothing would say so. A Long or Example that spells a region marker out in full is a hard error: the generated block would end inside itself.

--out defaults to <project>/.claude, resolved by the same walk everything else uses, and no marker is an answer rather than a fallback.

Keeping this file true

  • Scope: the shape of the command tree — the registry, the shared helpers, the three conventions, the round trip, the generator. The file table names every source file in this directory.
  • Update it when a command file is added or removed, a group is added, a shared helper in flags.go/sync.go changes, an exit code gains a meaning, a command gains a confirmation gate, the push/pull ordering guarantees change, or the rules around writing into .claude/ change.
  • Do not copy a flag list or a command's Long text here. kettle help <name> and the generated SKILL.md blocks are the two places that exist for it, and a third would be the one that drifts.