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>
The plugin required `tea`, Gitea's own CLI, for everything that is not an
issue: releases, pull requests, milestones, branches, actions, webhooks. That
put a second binary, a second set of logins nothing here could see, and 400
lines documenting somebody else's flags outside anything this repository can
test. One command over the transport that already existed removes all three.
Transport: `post` — the hand-rolled request the SDK cannot express, written for
the dependency endpoint — is generalized to an exported `Do`, and `post` is
three lines on top of it. Same http.Client, so the same RoundTripper files the
body under .kettle/payload/, the same `token …` header authenticates it, and a
non-2xx is the same *APIError. It does not paginate, does not reformat the
answer, and names no domain concept, so the layering test is untouched.
The endpoint rule is `tea api`'s, so an endpoint table written for that tool
still works — with one restriction it did not have: a full URL must be on this
instance. Every request carries the project's token in a header, and a URL on
another host would hand the token to whatever was typed.
Command: `kettle api <endpoint>` in a new `api` group, so the generator writes
plugins/kettle/skills/api/SKILL.md — group, directory and /kettle:api are one
word. No --repo and no --login, for the reason no sync command has them: a
cross-repository address is an address, and another instance is KETTLE_URL.
`-X DELETE` needs `--yes`; a flag typed on purpose is an operator's decision.
Scopes: a token minted for issues carries write:issue and answers 403 on the
first request outside issues, naming no scope. Gitea cannot be asked what a
token may do — its own token listing needs a password — so `auth add --scopes`
records it, `auth list` and `config` show it, and a 403 says which category it
is likely to be. Documentation only; nothing is checked against it.
skills/use — the tea reference, 239 lines of it — becomes skills/api: what to
ask for, which endpoints paginate, and how to write a body. Every mention of
`tea` as a requirement is gone from the manifests, the READMEs, the runner and
the four other skills; what survives is the back-compat with the old plugin,
which is a decision and not a debt.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
There is no CI: the instance has no act_runner and none is planned, so releases
are cut by hand. That makes `make check` the only thing standing between a
mistake and the tracker, and it is one command: gofmt, vet, the suite with the
cache defeated, `go mod verify`, a vendored build, and `kettle gen skills
--check`. The last one is the invariant worth having — the plugin's SKILL.md
command reference is generated from the binary's registry, so a flag that
changed cannot ship with documentation that recommends the old one.
`cli/cmd/release` publishes to Gitea using the same SDK the binary already
vendors, which is a pleasing thing to be able to say: nothing third-party
handles the artifacts. It is a second binary rather than a `kettle` subcommand
on purpose — `kettle`'s command tree is what generates the plugin's skills, so a
verb there ships to every operator, and publishing a release is build
infrastructure. It is idempotent end to end: an existing release for the tag is
reused, an asset of the same name is replaced rather than doubled, and a retried
run converges instead of duplicating.
`make release` refuses three things, each with its own message: a dirty working
tree, a TAG that is not what `git describe` reports, and a tag the remote does
not have. A release built from uncommitted code is unreproducible and nobody
finds out until they need to reproduce it.
`kettle version` reports the stamp, the toolchain and the VCS revision. The
default is `dev`, and a hand build says so and means it — a binary out of
somebody's working tree is not a release and must not claim to be one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
BREAKING: the plugin is `kettle`, not `tea`, and its commands are `/kettle:*`.
It also now needs a binary on PATH that it did not need before; the README and
every skill say how to get one and what a missing one looks like.
The plugin was 3800 lines of Python doing what a compiled binary does better,
and the name pointed at a tool that no longer takes part: `tea` is Gitea's CLI,
and since the transport moved into the binary nothing here shells out to it for
issues at all. A plugin named after it was going to keep suggesting otherwise.
Deleted: 19 scripts, the 14-file unittest suite, and the tea-guard hook. The
guard blocked any `tea` invocation that would run under a login the model picked
instead of the operator; the binary holds its own credentials and reads the
pinned login out of the project's own config, so that failure is no longer
expressible and there is nothing left to police. agents-sync stays — it is about
AGENTS.md symlinks and has nothing to do with any of this.
What the plugin keeps is what only a plugin can carry: the rules an operator
states and a binary cannot enforce. `init` still refuses to run inside a linked
worktree and still may not be model-invoked, because which directory is the
project is a statement a person makes. The issue format reference stays here and
stays the source of truth. The runner subagent is still for batches and still
may not decide what an issue says.
The command reference in the issue, sync and project skills is GENERATED from
the binary's own command registry, between markers, so a flag that changed
cannot ship with a skill that recommends the old one. `kettle gen skills
--check` exits non-zero when they drift. The generator owns the region and
nothing outside it: the frontmatter description, which is what decides whether a
skill loads at all, stays hand-written.
`use` survives and is the one place `tea` is still named — for releases,
webhooks and actions, which kettle does not cover. Its instruction to write
`--login "$GITEA_LOGIN"` and let the hook substitute the pin was true until this
commit and is now rewritten: `tea` keeps its own configuration, kettle keeps
its own, and configuring one configures nothing in the other.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
tea and tdl were two repositories, each carrying its own
.claude-plugin/marketplace.json — two marketplaces to register for what
is one collection. Fold them into one.
The repo root is now the marketplace and nothing else: a single
.claude-plugin/marketplace.json whose entries point at ./plugins/tea and
./plugins/tdl. A plugin's root is its own directory under plugins/, so
${CLAUDE_PLUGIN_ROOT} still resolves inside it and every path a plugin
uses stays relative to itself — the hooks and the test roots needed no
adjustment beyond the move.
tea's files move with git mv, so its history and blame follow. tdl
arrives as a plain copy; its history stays in claude-skills/threedotslab.
test_payload_root asserted `tmp/` was ignored by REPO/.gitignore. The
rule is that tmp/ is ignored, not which file says so, and git reads every
.gitignore on the way up — so the test now walks up to the repo root the
same way git does.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The plugin is issues and nothing else now. `skills/page` (the page-tree
domain) and `skills/wiki` (its bridge to a Gitea wiki) are gone, and
with them the issue domain's `wiki:` field — page titles were the only
thing that tied the two domains together, and a field the tracker has
no column for never came back from a pull anyway.
What is left is the shape AGENTS.md already claimed for the rest of the
repo: one domain, one bridge, one transport. The docs, the plugin
manifest, and tea-runner's skill table now say so too, and
test_payload_root walks the one script directory that remains.
Also removes openspec/config.yaml; nothing in the repo referenced it.
378 tests pass.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
README described a three-skill plugin that ships six. skills/page and
skills/wiki were absent from What it ships and from the project tree, so
/tea:page and /tea:wiki could not be discovered from the front page at
all; labels.py, close.py, evict.py, issue_evict.py, pin.py and the
agents-sync hook were missing from the tree too. Layout now matches
AGENTS.md, and says which of the two is authoritative.
The sync skill promised that push writes the computed branch back into
the issue file. It cannot: a successful push deletes the file, which
push.py:260-271 and its docstring already said. The paragraph now says
what happens instead — the ref goes up, and the branch comes back on the
next pull, from the tracker. The three claims around it were correct and
are kept.
Closes#27Closes#31
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
`_gitea.require_login` walked up from CWD and nowhere else. A worktree is a
sibling of the main checkout, not a descendant, and `settings.local.json` is
untracked — so the pin lives in the main checkout only, is not on the
worktree's parent chain, and the whole tracker half of the plugin died there
with "no login pinned". In the same directory the guard resolved it fine,
because it had a search of its own: one order, written twice, disagreeing.
It is written once now, in skills/auth/scripts/pin.py, and both callers import
it — the transport and hooks/tea-guard.sh. $CLAUDE_PROJECT_DIR, then a hint the
caller supplies (the hook passes its payload's cwd), then the current
directory; each searched up its parent chain, and only if that finds nothing,
across into the main working tree of a linked worktree met on the way, reached
by reading `gitdir:` out of the `.git` FILE and following `commondir`. No
subprocess — a PreToolUse hook runs before every Bash call and must not fork to
answer this.
The search still starts at the working directory and never at `__file__`,
deliberately asymmetric with `issue.store_root` and `_gitea.PAYLOAD_ROOT`.
Where an installation keeps its files is a fact about the installation; whose
login a project runs under is a fact about the project, and a plugin pointed at
somebody else's tree must not answer that from its own directory. pin.py says
so in as many words, so the next reader does not "fix" the inconsistency.
Two consequences fall out of it. `/tea:auth` no longer has any reason to run
inside a worktree, so no second pin lands in a directory that is deleted with
the branch — the skill now says to write it beside the common `.git`. And the
scripts can run where the work is: the workaround the bug forced, cwd in the
main checkout, made push.py send that checkout's branch as `ref`, which is the
one thing `branch:` exists to record.
tests/test_login_pin.py holds both halves: the hop against a hand-built layout
and against a real `git worktree add`, a run from the worktree finding the
login, no pin anywhere still erroring, the scripts' own directory not becoming
a source, `ref` coming out as the worktree's branch, and the hook and a script
answering the same directory alike. Two mechanical checks keep the callers from
growing a second copy of the walk. Three existing fixtures now copy
skills/auth/scripts, which the transport imports.
Refs #24.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Gitea becomes the source of truth. Once a push is confirmed, push.py
deletes tmp/issues/<id>.md and <id>.comments.md and prints the number and
URL the issue now lives at; the current state is obtained by pulling
again rather than by reconciling. --update follows the same rule, with no
exception: what is local is what has not left.
This reverses three statements AGENTS.md used to make, and rewriting them
is part of the change:
- "tmp/issues/ is the store, not a cache of Gitea" — it is both, split
by origin:. An origin: local file is the only copy of the work; an
origin: gitea file is a deletable working copy.
- "Pushing is additive: the file is never deleted" — it is deleted.
- "origin: local is a durable state" — complete, but not durable:
pushing ends it.
Slug stability, which the format promises for the life of an issue, can
no longer rest on a file push is about to delete. The slug goes up in the
body as a hidden marker, <!-- tea:id <slug> -->, on the first line:
map.to_payload strips every marker and prepends exactly one, map.from_api
strips every marker on the way down, so the local file never holds one
and a body cannot accumulate them however many round trips it makes. The
marker survives a rename in the web UI, a lost .remote.json, a fresh
clone and another machine — none of which a local index does.
Deletion is the last thing that happens to an issue and only after the
transport returned, the answer carried a positive integer number (and, on
--update, the number that was PATCHed — push.confirmed_number), and
.remote.json was written. A raised transport, a non-2xx, an empty or
mismatched body each leave the file on disk and stop the run.
.remote.json is no longer "only an index over the files": its entries now
deliberately outlive them, so it is the local number -> slug ledger and
rebuild_map merges into it instead of reconstructing it from files that
may be gone. It stays recoverable, from the markers in Gitea rather than
from the files. push.dep_state reads it too, so a blocker whose file an
earlier push dropped still gets its native dependency link.
Also fixes a pre-existing bug the new tests hit: issue.all_ids treated
<id>.comments.md as an issue called "<id>.comments", so a bare push.py in
a store holding pulled threads tried to file a comment thread as a unit
of work. A slug has no dot in it.
tests/test_drop_after_push.py covers the round trip (push -> gone -> pull
-> identical in slug, depends: and body), the marker's algebra, and every
failure path separately. test_push_dependencies.py is updated where it
encoded the old "never deleted" contract. 183 tests, no network.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
A checkbox is the one part of a body that is state and not prose.
Everything else is written once; boxes get ticked as the work goes, and
until now the only ways to tick one were a human with an editor or a
model rewriting the whole body. The second is worse: the rewrite re-flows
lines and re-words sentences, so the issue's diff swells around a change
that means one character. Progress was invisible too — issue_index.py
builds INDEX.md from metadata and never looked inside a body, so "3 of 7
done" required opening the file.
All three pieces are domain: a checkbox is body syntax, which is part of
the answer to "what is an issue". The parser goes in issue.py so the sync
layer can reuse it instead of redefining the format on its own side.
issue.py gains checkboxes(text) -> [Checkbox(index, line, end_line,
checked, text, section)], plus set_checkbox(text, item, checked) and
checkbox_progress(text). All pure, no I/O, importable from another layer.
The scan covers the whole text, in any section: the type/feature template
keeps child issues as checkboxes under `## Issues`, so binding the parser
to `## Acceptance criteria` would silently lose half of them; the heading
is recorded, never required. Only a marker line opens an item, so a
wrapped continuation line belongs to the item above it rather than
counting as one of its own. A `- [ ]` inside a code fence is an example
of the markup and is skipped. Line numbers are relative to the text
given, which is what lets a caller work on a body or on a whole file.
issue_ac.py lists the items numbered, grouped by heading, and ticks one
by number or by substring. An ambiguous substring is an error that prints
the matches — a coin flip would tick the wrong box and look like it
worked. It patches the file rather than round-tripping through
Issue.to_text(), so exactly one character changes: metadata order,
wording, wrapping, trailing whitespace and CRLF endings all come back
byte for byte, proven by a diff in the tests.
INDEX.md gains a progress column: `3/7` for an issue with checkboxes,
blank for one without. Counted off the body at build time and stored in
no field — a second copy of the state would be wrong by the next edit.
issue_check.py is unchanged and stays that way on purpose: an unticked
box is work not done yet, not a malformed issue, and validate() carries a
comment saying so.
Delivering a tick to the tracker is out of scope — that is push.py
--update in /tea:sync.
format.md gets one clarifying bullet. It said acceptance criteria are
checkboxes but never said what a checkbox is, so the parser had to settle
questions the format left open: any section, wrapped items, fenced
examples. Those rules are now written down where the parser and the sync
layer can both point at them.
tests/ is new, and is the convention: plain stdlib unittest, no pytest
and no third-party deps, since the code under test may not have
dependencies either. Scripts are imported via sys.path.insert and every
fixture is built in a TemporaryDirectory, never in tmp/.
python3 -m unittest discover -s tests -v 32 tests, OK
skills/issue/scripts/ still imports stdlib only, with no subprocess.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The skills carry meaning, the scripts carry work. Splitting the second
half onto a cheap model keeps the main session's context for the part
that needs judgement.
tea-runner is a Haiku subagent with Bash/Read/Grep/Glob/Skill and
nothing else. It loads /tea:sync or /tea:issue for the command table
rather than carrying its own copy, so the skills stay the single source
of truth for the script surface.
It executes and reports; it decides nothing. No Edit and no Write, so an
issue body is out of reach. No raw tea, no --force, no closing or
retitling, no pushing past the set it was handed, one retry maximum. A
failed validation, a missing type, an unpushed dependency come back as a
question in a `blocked:` line. The reply is a fixed receipt — commands
with ok/FAIL, touched paths, stderr verbatim — never a payload dump.
Knowledge still flows one way: nothing under skills/ knows the agent
exists, and deleting agents/ changes nothing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An issue was a Gitea row that happened to be cached locally: its identity
was the tracker's number (42.md), its dependencies were tracker numbers
(depends: [#12]), and a local issue existed only as a draft that push
deleted on success. Nothing could be planned or tracked without a tracker.
Split into layers, with knowledge flowing one way:
skills/issue DOMAIN what an issue is: format, validation, dep graph
^ offline; stdlib imports only, no subprocess
| imports
skills/sync BRIDGE map.py md <-> Gitea JSON, pure, no I/O
_gitea.py login pin, api, pagination, filters
skills/use REFERENCE tea CLI docs for non-issue entities
skills/issue never imports skills/sync. Delete the sync layer and the
domain keeps working.
Identity is now a slug derived from the title (wire-sqlc-appclick.md) and
is stable across retitles and pushes. Tracker numbers live in a `gitea:`
field, never in a file name and never in `depends:`; the pair is indexed
in .remote.json, which is a cache over the files, not a second source of
truth.
Behavior changes:
- Pushing is additive. The file is never deleted; it gains gitea:/url:/
synced: and origin: flips from local to gitea. `origin: local` is a
durable state, not a pending one.
- Pushes go in topological order so dependencies get numbers first.
- The dependency graph is computed offline from `depends:` metadata; body
prose is passed through unchanged in both directions rather than being
rewritten between slugs and #N.
- `origin` is domain-owned (whether work exists elsewhere is a fact about
the work); the handle and how to reach it stay with sync.
Script moves:
issue_get.py -> sync/pull.py
issue_push.py -> sync/push.py
issue_list.py -> sync/remote.py
issue_index.py -> issue/issue_index.py
_tea.py -> split into issue/issue.py, sync/map.py, sync/_gitea.py
New: issue/issue_new.py, issue/issue_check.py, issue/issue_tree.py, and
sync/comment.py — comment posting was the last issue operation still
hand-rolled through raw `tea api`.
references/issue-format.md moves to skills/issue/references/format.md;
label hex colors move out of it into map.py, since a color is how a
tracker paints a chip, not what an issue is.
Verified: offline path end to end (new, check, tree, index, push
--dry-run) and read-only against Gitea (remote listing, pull with
mapping, comment guard). Write paths of push.py and comment.py are not
exercised here.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Replace fetch_issue.py with four scripts around a flat, greppable cache in
tmp/issues/. Planning stays offline and issues reach Gitea in one push:
- issue_get.py: fetch by key or by filter (--milestone/--label/-q). The list
endpoint carries issue bodies, so a whole milestone costs one request per 50
issues. Gitea silently ignores an unresolvable milestones= filter and returns
the entire backlog, so the milestone is resolved up front and every returned
issue is re-checked locally. --deps walks the dependency graph downwards via
the structured sections plus native dependencies and writes tree-<slug>.md.
- issue_push.py: validate a local draft against the canonical format, create
missing labels with the right colors and exclusivity, POST, delete the draft.
- issue_list.py: discovery to stdout, writes nothing.
- issue_index.py: rebuild INDEX.md from what is on disk.
Files use one metadata field per line with inline lists so plain grep works
without a parser. This is a cache and a drafting area, not a mirror: no drift
tracking, no sync back.
Projects are not fetchable — the projects API is 404 on Gitea 1.26; documented
alongside the milestone caveat.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`/tea:login` could not be invoked as `/login` because Claude Code
reserves that name as a built-in command. Renaming the skill to `auth`
frees the short form: `/tea:auth` and `/auth` both resolve.
Updates all references in the guard hook, README, plugin manifests, and
the use skill.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The install steps used a non-existent "plugins" settings key, so a fresh
clone loaded nothing. Add .claude-plugin/marketplace.json and document the
supported /plugin marketplace add + /plugin install tea@tea flow.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
naudachu
naudachu