The transport was hand-rolled net/http against the REST API. The payload shapes
were ours, in internal/wire, which meant every field Gitea learned was a field
somebody here had to notice; and "does this instance have issue dependencies?"
had to be guessed from a status code, because a 404 from a missing route and a
404 from a missing issue look alike.
The SDK settles both. The shapes are maintained by the people who maintain the
server, and the client negotiates the server's version when it is built, so the
dependency endpoint is now gated on `>= 1.20.0` — verified against the release
where the route appears, not assumed. Below the gate nothing is requested at all.
internal/wire keeps what the SDK has no answer for: addressing. The SDK takes an
owner, a name and an int64 and never parses, while `42`, `#42`, `owner/repo#42`
and an issue URL are four spellings of one address, all four are what somebody
has in hand, and Key is what the ledger is keyed by. The payload structs go.
Four things that had to survive the move, and did:
- request bodies still land in .kettle/payload/, now via an http.RoundTripper on
the client the SDK is given — which is better than before, because it files
every request rather than the ones a call site remembered to name;
- errors still carry the HTTP status AND the response body, and a decode failure
on a 2xx is deliberately not an APIError, so the dependency probe cannot read
a bad decode as "feature missing";
- the number -> slug ledger is untouched, entries still outlive the files they
name;
- Client.For(repo) still re-points at another repository for one call.
What it cost, written down in AGENTS.md where it happened. internal/mapping's
layering test was a fact about the import graph — nothing in its closure could
open a socket — and the SDK ships its types and its client in one package, so
the test now asserts what is still true: the bridge performs no I/O, checked on
direct imports plus a grep for time.Now. A run makes one extra request before it
does anything. Gitea's issue edit endpoint carries no labels, so a push whose
labels changed needs a second call; push makes it and says so. go.mod requires
go 1.26, which the SDK sets and which is now the floor for building this binary.
internal/issue and internal/project are byte-identical. The domain did not
notice, which is the whole argument for the layering.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The plugin resolved its issue store from `__file__`, which put it inside a
versioned plugin cache: issues written from one project were invisible from the
next, and `origin: local` files — the only copy of that work by definition —
were stranded a version bump at a time. The walk that answers "which directory
is the project" was written three times over, and in a linked worktree the three
disagreed. Both are runtime failures rather than logic ones, so the fix is a
compiled binary: one walk, imported rather than re-derived, and a layering rule
the build graph enforces instead of a grep.
Seven packages, knowledge flowing one way. `project` answers which directory is
the project and depends on nothing. `issue` is the domain — format, taxonomy,
validation, checkboxes, dependency graph, the store, eviction — offline, with no
tracker in it. `wire` holds the protocol shapes. `gitea` is the transport,
`mapping` the bridge, `config` the credentials, `cmd` the command tree. Four
tests hold the boundaries, each failing on a real mistake rather than a naming
convention.
The marker moves to `.kettle/` and the login pin moves out of the harness's
settings file into `.kettle/config.yaml`, which pins a login by NAME; the tokens
live in one file per machine, mode 0600, outside every working tree. That
retires the PreToolUse guard hook entirely — the binary holds its own
credentials, so a command running under a login nobody chose is not expressible
rather than caught.
`kettle init` migrates an older `tmp/issues` or `.tea/issues` store in, as a
move: a store left behind at an old path is one somebody edits by accident
months later.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
naudachu