All the pieces that occurred on github.com/Cocoanetics because the SwiftCross put up (June 2 – June 13, 2026). 4 new kits, a fourteen-year-old flagship that got here residence, and the stretch the place the extraction sample began delivery to customers.
Once I printed the SwiftCross put up on June 2nd, I left two threads dangling: an agent was refactoring SwiftMCP’s package deal construction so client-only customers wouldn’t drag in SwiftNIO, and if that labored it will unlock SwiftAgents on Home windows. I anticipated to report on that “finally.” Each landed earlier than the day was over. That set the tone for the week — I preserve discovering that the limiting issue is not how briskly issues get constructed, however how briskly I can resolve what to construct subsequent.
The free ends didn’t keep free for a day
SwiftCross itself didn’t sit nonetheless previous lunch. By the night of June 2nd it had grown WallClock — a conveyable nanosecond wall-clock (clock_gettime on POSIX, GetSystemTimePreciseAsFileTime on Home windows) that I wanted for span timing in agent traces — launched as 1.1.0, and Setting.set, a conveyable setenv for .env loaders, launched as 1.2.0. Two tags in a single night for a package deal I had introduced that morning.
The identical night, SwiftMCP 1.5.0 shipped the package-traits refactor from the put up: Server, Consumer, and OpenAPI traits, all on by default, with swift-nio, swift-crypto, and swift-certificates gated behind Server. A client-only shopper now will get a completely NIO-free construct — which is the entire ballgame for Home windows. (1.5.1 adopted the subsequent morning to widen the swift-crypto vary; dependency hygiene by no means sleeps.)
And SwiftAgents cashed the test instantly: it now hyperlinks SwiftMCP with traits: [Client] solely, dropped swift-nio totally, migrated its remaining shims to SwiftCross, and the Home windows CI probe went inexperienced. The factor I described as “if that works out” within the final put up was a merged PR about three hours later.
SwiftMCP then bought two extra modernization passes in the course of the week: all three server transports now conform to swift-service-lifecycle‘s Service (swish SIGTERM/SIGINT shutdown replaces my bespoke sign handler, and the stdio transport lastly exits on EOF as a substitute of politely polling at 10 Hz), and the HTTP transport was rebuilt on apple/swift-http-types — typed header fields and strategies as a substitute of strings, with NIOHTTPTypesHTTP1 deleting a pile of hand-rolled conversions.
The cascade reached SwiftMail
That is my favourite a part of having a small ecosystem: a launch in a single package deal turns into a one-line diff within the subsequent. SwiftMail 1.7.0 (June 3) adopted SwiftCross 1.2.0, dropped my swift-nio-imap fork in favor of Apple’s upstream, refreshed each dependency, and bought Home windows and Android CI. A day later 1.7.1 mounted a genuinely refined bug — and never considered one of mine: contributor tabmail-kmyi discovered that the pipelined FETCH dispatcher may drop later physique components when a server batches untagged responses. Mounted with correct EmbeddedChannel regression assessments. The packages being presentable sufficient that strangers debug IMAP pipelining for you is a quiet milestone of its personal.
The brokers audited their very own playground
On June 4th I did one thing that also makes me smile: I let an agent free inside the SwiftBash sandbox and instructed it to poke round like a suspicious consumer. It got here again with a stack of points — stat leaked my actual host uid/gid by the digital filesystem, the time key phrase didn’t parse, ps couldn’t see the shell it was operating in, /tmp didn’t present up in ls /, assorted GNU choices have been lacking from xxd, od, and realpath.
Then, over that night time and the subsequent day, the brokers mounted every part on their very own checklist. SwiftBash now experiences digital identification on the filesystem boundary, has an actual mount command with a digital mount desk, parses time and experiences actual/consumer/sys, and its ps is not blind to itself. ShellKit bought the matching ProcessTable lifecycle work, and BinCatalog moved from ShellKit into SwiftBash the place it belongs. The README gained a “What’s digital (and what which means)” part, as a result of if an agent wanted to find these solutions empirically, so would you. There’s one thing pleasingly self-referential about an agent submitting bugs in opposition to the sandbox constructed to include brokers — after which resolving them.
sqlite3 to byte-parity, then out the door
Additionally on June 4th, the SwiftPorts sqlite3 shell port went by a parity marathon: .dump constancy with correct identifier quoting, byte-exact floating-point formatting through a small C shim, .databases/.schema/.present/.width/.restrict conduct matched in opposition to the true factor.
It turned out that it was exceptionally simple to get sqlite3 to construct for all platforms as a result of the code is basically a single c file. After which I remembered that there was a full textual content search extension (FTS) that Apple selected to not ship in sqlite3. I requested the brokers, if we may that. They sayed “Sure!” and constructed it. Then I remembered that there was additionally a vector search extension for sqlite3. Identical story: if you happen to can assume it, they will make it.
For this reason we bought FTS5 and sqlite-vec which you’ll be able to allow individually through SPM traits.
At which level the SQLite stack had clearly outgrown its host, and on June eighth it was extracted into SQLiteKit — a brand new repo: a cross-platform Swift wrapper over a vendored SQLite amalgamation with ready statements, FTS5 full-text search, and sqlite-vec semantic search behind traits. Twelve thousand strains left SwiftPorts in a single commit. If that transfer sounds acquainted, it’s precisely what SwiftCross was: the third time you copy one thing, it needs to be a package deal.
Instructing the brokers to look
SQLiteKit instantly had a buyer. For a very long time I saved saying that macOS and iOS must get an un-neutered model of sqlite3. Since final 12 months CoreSpotlight clearly has a semantic index, which cries out “vector DB”. However there is no such thing as a first-party public semantic DB functionality to be discovered on Apple platforms.
Final 12 months I experimented with a easy linear vector search in SwiftAgents, in addition to the semantic search that OpenAI offers us. There you add paperwork to OpenAI and you should use a file search software to complement your completions.
Now it was inside my grasp to develop SwiftAgents a persistent SQLiteVectorStore — vec0 KNN plus FTS5 bm25 plus hybrid search with line-span provenance — after which, in a single day of eight consecutive PRs, absorbed the retrieval methods from tobi’s qmd: sensible markdown chunking that respects code fences, reciprocal rank fusion, lex/vec/hyde question enlargement, LLM-gated enlargement, a batch reranker, and question/doc embedding asymmetry. On-device Apple NL embeddings by default; OpenAI solely when a secret’s current.
The deliberate half was the order. qmd — question markdown — and the OpenClaw equal every carried numerous hard-won, embedded data about tips on how to construct a memory-retrieval pipeline, and I didn’t need that data trapped inside a CLI. So I had the agent rebuild the index engine inside SwiftAgents first, the place something can reuse it, and solely then — as a result of the engine existed — QMDKit bought constructed on prime of it in about two days: one other new repo, a qmd clone that indexes and semantically searches Markdown, accessible as a standalone CLI and as a sandboxed ShellKit builtin. Which suggests SwiftBash classes — the identical sandbox the brokers work in — now have semantic search over my notes as a shell command. The brokers constructed their very own retrieval tooling, then put in it in their very own atmosphere. I simply watched.
GitKit, and a lesson in humility from libgit2
June ninth: I needed SwiftPorts’ git to cease relying on a private libgit2 fork. The fork I’d been leaning on minted a recent department and package deal for each single libgit2 launch; I needed the alternative form — one construct that pulls the C supply straight from upstream as a git submodule, version-matched, so bumping libgit2 turns into a submodule pointer slightly than a brand-new repo. So GitKit grew to become the third new repo of the week — libgit2 1.9.4 packaged for SwiftPM with a pristine upstream submodule, pushed to inexperienced on macOS, iOS, Linux, Home windows, and Android in a single afternoon (Home windows wanted a curated header umbrella; Android wanted its personal toolchain rituals; tvOS and watchOS have been dropped as a result of libgit2 forks processes, they usually don’t).
Then actuality arrived on schedule. SwiftPorts adopted GitKit, and git stash apply broke. The perpetrator was splendidly obscure: feature-define names inherited from one other fork that not matched libgit2 1.9.4’s dialect, silently disabling issues like nanosecond timestamps. Undertake, regress, revert, repair, re-adopt — all inside about 9 hours, each leg verified by the five-platform CI. That is the outer loop from the SwiftCross put up doing precisely what I claimed it does, besides this time the pink checkmark saved me from delivery a subtly damaged git.
The week closed with the now-familiar transfer: SwiftPorts’ complete Swift git layer was lifted out and have become the GitKit 2.0.0 SDK — a full idiomatic Repository API over libgit2 — leaving SwiftPorts with simply the sandbox-aware face and CLI. Zero duplication. (There was additionally a short comedy in three acts about how a tagged package deal might rely upon the untagged swift-archive; the reply, after one rejected fork-with-a-minted-tag, is: pin the upstream commit and let traits prune it.)
The parser needed to be a package too
If the primary 9 days had a chorus, it was “the third time you copy one thing, it needs to be a package deal.” Day eleven onward simply saved proving it — beginning, of all locations, with the oldest piece of code in the entire org.
For years the tolerant HTML parser inside SwiftText was a quiet workhorse: a Swift rewrite of DTHTMLParser, itself a chunk of DTFoundation that predates most of those repos by greater than a decade. On June 12 it lastly moved out by itself. XMLKit — an async streaming HTMLParser constructed on libxml2’s SAX interface, so it handles real-world malformed HTML gracefully and by no means builds a DOM except you ask it to. It left SwiftText carrying its full git historical past, and SwiftText instantly adopted it again as a dependency. Launched as 1.0.0 that afternoon, then 1.0.1 a number of hours later as soon as CI was constructing and operating the check suite on all 5 platforms — macOS, iOS, Linux, Home windows, Android. The identify is a promise: an XML SAX module alongside the identical strains is the apparent subsequent tenant.
The trustworthy set off was narrower than the outcome. I didn’t wish to drag swift-markdown into DTCoreText simply to achieve the HTML parser, and my first intuition was a budget repair — a trait inside SwiftText that gates the dependency away. However the earlier extractions had taught me higher: the third time you attain for a factor by a wall of unrelated dependencies, it needs to be its personal package deal. XMLKit is, at coronary heart, libxml2 plus a skinny layer of Swift wrappers — and as a result of it stands alone, with nothing else cluttering the manifest, I may chase all of it the best way to inexperienced CI on all 5 platforms. That five-platform checkmark is my present gold normal, and realistically you solely earn it one library at a time.
After which the fourteen-year-old flagship got here residence
That is the half I didn’t see coming. DTCoreText — the CoreText-meets-HTML library with six thousand stars and a decade and a half of historical past — shipped 2.1.0 on June 13, and it’s probably the most fascinating launch of the stretch exactly as a result of it’s the least new.
Two issues occurred to it without delay. It bought options: HTML tables with an actual NSTextTable-compatible mannequin (colspan/rowspan, border-collapse, row-by-row pagination), CSS floats (float/clear with textual content wrapping), CSS border-spacing, and — the one which made me grin — native iOS 27 TextKit bridging, so DTCoreText’s textual content blocks and tables lay out as actual NSTextBlock/NSTextTable on the brand new SDK. Roughly 8,600 strains throughout 5 PRs, the primary function launch constructed on this spring’s Swift 2.0 rewrite.
The tables particularly are a fourteen-year lastly. The primary request to render an HTML
