Changelog¶
All notable changes to the SDK (Loopl, LooplRuntimes, LooplUI), the app and this site.
Versions follow SemVer; dates are UTC. Pre-1.0, minor versions may break the API — each
breaking change is called out under Changed with the old → new symbol.
Pre-release
loopl is in its first week. The SDK surface is settling; until 0.2.0
ships, treat the pages under SDK as the direction, and the changelog as the contract.
Unreleased — human in the loop & http¶
- A test stopped failing for the first three hours of every day.
ContactsPhotosToolsTestsasked the photos tool for pictures taken today and required at least one back, while its fixture creates pictures taken one, two and three hours ago. Those are the same day for twenty-one hours and different days for three, so between 00:00 and 03:00 in the runner's timezone the tool correctly returned nothing and the suite called it a failure. Nothing was ever wrong with the product: the same commit passed at 23:47Z and failed at 00:08Z. The test now asks for the day its own fixture used, which is the behaviour it meant to check — filtering by a given day returns that day's pictures. Proven by ageing a fixture past midnight on purpose: asking for "today" finds nothing (the old shape fails) and asking for the fixture's day finds them. - Four more messages about your own models stopped naming file formats at you. When the Mac and PC build of a model you trained failed, the app said "GGUF:" followed by a raw error — even though the line directly above it said "iPhone conversion failed" and the same screen's own estimate calls those two things the iPhone version and the Mac and PC version. It says that now. The sentence telling you where your trained model went listed three private repos with weights and GGUF in brackets; it now says it saved them privately to your Hugging Face account, and what each one is for. Models › Mine, when it has nothing to show, used to say "No MLX or GGUF repos under @you yet" — which names two runtimes and answers nothing; someone looking at an empty list under their own name wants to know why their other models are missing, so it now says loopl has nothing it can run there yet. And the note under a model's file list stopped listing weights, tokenizer and configs and now says simply that only the files needed to run the model are downloaded, not its readme or page history.
- When a model will not start, loopl now says what is wrong and what you can actually do. The message had been telling people to tap Repair — and there is no Repair button in loopl, and never has been: the repair happens by itself, and the only thing to tap is Try again. At one point it managed to say "tap Repair" and "Repair failed" in the same breath. It had also been one message doing duty for three different situations, which meant it was wrong in two of them. Now each says its own thing: if you are offline it tells you it can fix this with a few kilobytes once you are connected, and that the model itself stays on your device; if the fetch failed it says so and Try again is worth a tap; and if the model's files are already correct and it still will not run, it says plainly that nothing here will fix it, that this is in how the model was published rather than anything you did, and suggests another model — without dangling a retry that cannot work.
- Four other messages about models stopped using the vocabulary of the people who package them. Tokenizer files, the weights, a few KB, checkpoints, gated repos, conversions, Metal and the quant all described things in words that only make sense if you build models for a living. The line under the Hugging Face results was the worst of them, because it was also untrue: it claimed only "chat and instruct checkpoints" were listed, when what is actually left out is models made for speech, transcription and search — a plain base model was always listed. That line is the only place the app explains why a model you expected is missing, so it now says exactly which kinds are left out, and that models needing their owner's permission are among them. The Format picker stopped describing itself and started saying what the choice means, including that opening a GGUF model is where you pick a size: smaller is faster, larger answers better. The names MLX and GGUF stay, because someone choosing between two formats needs them — a word is only jargon when the reader cannot act on it.
- History's delete flow is now verified on an iPad, and the four tests that said it was broken were wrong four different ways. The full UI suite had never run on a tablet until this week, and when it did, all four
HistoryUITestsfailed on the screen that deletes someone's conversations — withkAXErrorCannotComplete … kAXScrollToVisibleAction, a message that names a scroll and reads like a flake. The element it could not scroll to was 36 pt wide and 0 pt tall at{417, 611}: the exact centre of the iPad, which is where the invisible ⌘Y proxy button lives.keyboardShortcutkeeps that proxy in the accessibility tree even through.accessibilityHidden(true), it carries the label History like the real control, andapp.buttons["History"].firstMatchhad been picking it — so every one of the four tests died on a button nobody can see, before reaching the screen under test. Behind it were three more phone-shaped assumptions: that the iPad's sidebar toggle is ours (it is Apple's, and ships with no identifier at all), that the navigation bar states “3 selected” (the sidebar's bar hasCancelandSelect allleading and the toggle trailing, so UIKit drops the title — the count survives on the Delete 3 button, which is the one that matters), and that History always offers Select (on an empty archive, offering nothing is the correct behaviour being tested). A sixth failure in the same family blamed ⌘Y on every iPad while probing for a sidebar section header that does not exist; with a probe that can actually succeed, ⌘Y works. Nothing in History was broken on an iPad — select, select all, the count, the confirmation, the delete, the empty state and Cancel all behave — but nothing had been proving it. One sharedreachHistory()now opens History correctly on either shape, and a failure prints the buttons that were on screen instead of the word scroll. - A documentation change stopped spending an hour of build machines.
cianduieach carry a list of paths that should not start them, so editing the docs does not occupy a macOS runner — and the previous change (the changelog fragments above) added two docs-only files that were in neither list: the new docs pull-request workflow and the release script that folds fragments together. Merging it therefore started a fullcirun and a two-deviceuirun to verify a changelog: about 80 minutes of build machines for zero bytes of app. The minutes are not the real cost —uikeeps only one queued run at a time, so a pointless run can take the slot the next real app change needed, which is the same starvation that once left that guard with a single green run in its first fourteen. Both lists now cover the new paths, and the set lives in one place (scripts/check-workflows.sh, whichcialready runs) with the two workflows checked against it, so forgetting an entry is a red build instead of a silent hour. Checked by breaking it on purpose three ways — removing the two new entries names exactly those two, deleting a whole list names all six, a malformed workflow is still caught — and it immediately caught a slip of my own during the test, when restoring one file from git quietly took my fix with it. - A pull request no longer edits the one line every other pull request edits. Every entry on this page was added directly under
## Unreleased, which made that heading the only thing two concurrent branches both had to touch: one lane hit 6 conflicts in 6 pull requests, every one on this file and on nothing else, and each cost a resolution, a force-push and a full CI re-run on the one file in the diff nobody reads while reviewing. The real risk was not the minutes — a conflict resolved in a hurry is how a<<<<<<<marker reaches a merge commit, which is why merging has needed a grep for markers as a guard. A branch now writesdocs/changelog.d/<pr>-<slug>.md, a file no other branch has, and a mkdocs hook splices the fragments into the Unreleased section as the site builds, newest pull-request number first; the page a reader sees is unchanged, and the fragment files are dropped from the built site so they never become stray pages.scripts/changelog-release.pyfolds them into this file under a dated version heading at release time, which is the only moment anybody edits it. The hook carries a dependency-free self-test that runs in the docs workflow before the strict build, and it was checked against four deliberately broken versions of itself — ascending order, splicing below the existing bullets, a missing heading passing silently, and the directory's README published as an entry — each of which it caught. - A docs change is now proven before it reaches
main, not after. Nothing ever built the site on a pull request:docs.ymlruns on pushes tomainonly, because it is a deploy pipeline — it holdspages: write, sits in thepagesconcurrency group with cancel-in-progress and ends in a deployment, so simply adding pull requests to it would let a branch cancelmain's in-flight deploy and then try to publish itself. A separate deploy-freedocs-prworkflow now builds the site strictly on any pull request that touchesdocs/,mkdocs.ymlor either docs workflow, with read-only permissions and a per-branch concurrency group; social cards and the DocC reference stay onmain, which still builds the real site. This was not a hypothetical gap: the fragment mechanism above went through review with a landmine in it, and only a hand-run of the adjacent script found it — a fragment naming a test suite, which nearly every real changelog entry does, turneddocs/check_symbols.shred, because that script skipschangelog.md(release notes legitimately name things that live inTests/, notSources/) and knew nothing of the new directory. The first lane to write a normal entry would have reddened the docs job onmainand stopped the site updating.changelog.d/is now skipped the same way — and only for the symbol scan, so the copy rules still hold there. Checked by watching it fail and then pass: a fragment namingEngineArchiveTestsfails before the fix and passes after, a bogus type on a real docs page still fails, and a fragment using the product name the docs copy rule bans still fails that rule. -
Switching models no longer needs room for both in memory (build 43). Picking a second model used to load it while the first was still resident, so the peak of every switch was both models at once — and iOS kills an app at a memory ceiling (about 5 GB on an 8 GB iPhone) even when either model alone fits. That is why a switch between two models that each ran fine could end the app without a word. The current model is now let go first, then the next one is read from disk; your downloads are the library, memory holds one. If the new model fails to load, the previous one comes back from disk and the line under the chat says what went wrong; if that reload fails too, the banner says both. A switch while the model is still answering is refused, the same way the other buttons that would disturb an answer are, with a line that says to wait and pick again.
EngineSwitchOrderTestspins the order (unload, load, and on failure reload), the refusal, and the two failure messages. -
The accessibility-squeeze guard stopped crying wolf on an iPad. The check flags a sentence that has been given less than 45 % of the screen — which is right on a phone and wrong on a wide screen, because text has a maximum readable measure: on an 834 pt iPad a perfectly formatted chat bubble is 324 pt, a History subtitle 263 pt and a tool row's note 357 pt, and all three were reported as bugs. The real squeezes the guard was built from are far smaller in absolute terms — 90, 132, 139, 150, 163, 175 pt — so the rule is now
min(45 % of the screen, 220 pt), which separates the two sets cleanly. On every device the suite runs in CI the proportional term is still the smaller one (181 pt on a 16 Pro, 169 pt on an SE), so the guard is unchanged there; only screens wider than ~489 pt are affected. An availability chip, whose spoken label is long while its capsule is deliberately small, is excluded by name. - The on-device tool rows stopped speaking developer. Nineteen lines in Settings ’ Tools, each fine on its own, together carrying our vocabulary into a screen where somebody decides what their phone may do:
translatetold a person holding an iPhone it was “not available on Simulator”;open_appoffered a “deeplink”;describe_imagenamed “the Vision framework”;vibratebuzzed “the Taptic Engine in a few named patterns”;musicdrove “the system music player”;flashlightswitched “the camera LED”. They now say what happens — “Opens another app — Maps, Mail, a website”, “Names what is in a picture — scenes, objects, faces, QR codes — using your iPhone itself”, “Makes this iPhone buzz”, “Turns the flashlight on and off” — and two rows that said “on this device” while their neighbours said “your iPhone” simply stopped naming the device, since the section's own footer already says everything runs here (and a tools module cannot know whether it is an iPhone or an iPad). A test now reads all nineteen as a set on every push: no developer vocabulary, 140 characters, ends in a full stop, one name for the phone. - The app's most repeated layout bug now has a machine watching for it. Three separate fixes this week were the same shape: a long sentence and a short control sharing one row, so the sentence got whatever width the control left it. Settings ’ Tools (“Asks before POST, PUT, PATCH or DELETE” in a 139 pt column), the approval card's header, and the new “newer version” line (“Hug-ging Face has a…” in ~90 pt beside Get it). It is invisible at the text size a developer works at and obvious at Accessibility XL, which is why it kept coming back.
TypeSqueezeUITestsnow walks six screens at Accessibility XL and fails when a label of 25 characters or more has been given less than 45 % of the screen — it reports the sentence and the width, and it was checked against the old code: restore the pre-fix Settings row and it names four of them with their exact widths. - A model's name and its description get the full width of the row at accessibility sizes. In the loopl collection and “More models”, the curator's note shared its row with a monospaced size (1.75 GB), which at three times the type is wider than the note: “The best balance of speed and quality on iPhone” came out 150 pt wide on an SE. The size and fit marks now sit under the text, and the NAME — which had been truncating to “loopl…” next to its “sees pictures” badge — gets its own line. Nothing moves at normal text sizes.
- The app can finally tell you the model on your iPhone is out of date. Every catalog row pins no revision, so a download takes whatever the repo's
mainpointed at that day — and on 7 October all three loopl sizes were replaced in place by v1.2. Someone installing today got v1.2; someone who downloaded yesterday kept v1.1; the app said nothing either way, and there was no word for outdated anywhere in it. Models now records the commit each copy arrived at, asks the Hub what the repo serves when you open the tab (one cached request per model per day, nothing at all in offline mode), and when they differ the row reads “Hugging Face has a newer version. Your copy arrived on 28 September.” with a Get it button. Three deliberate silences: a copy that matches, a copy from before this release (loopl has no record of what it is, and inventing one would be telling you to spend 1.75 GB on a guess) and every row the Hub has not answered for. Nothing downloads without the tap — that is the whole reason this is a line and a button rather than a background update. - A pill wider than the card no longer hangs off the edge. Onboarding's model cards lay their badges out with a tiny flow layout that puts each badge at its ideal size and starts a new row when the width runs out — but a badge wider than a WHOLE row has no row to move to, so it was placed at its ideal size and simply ran off the card: at Accessibility XL, Recommended · fast + smart ended 15 points past the right edge of an iPhone 16 Pro. Such a badge is now re-proposed at the row's width and wraps inside its own capsule. Same layout, nothing moves at normal text sizes.
- Nine screens are now measured at an accessibility text size on every push. The overflow above, and the 502-point-wide chat before it, were both invisible to the test suite: nothing checked that the app stays INSIDE the screen when the type is three times bigger.
OverflowAuditUITestslaunches nine of them — chat with tool rows, an approval card, an ask_user form, render cards, Models, Settings with Tools open, Train, History, onboarding's pick page — under Accessibility XL and fails with the element and its coordinates, so a failure answers the question instead of posing it. Two exclusions, both measured rather than assumed: the system scroll indicator, which lives in the edge inset by design, and unnamedListcontainers, which report their cell background 16 points outside the window on every screen at every text size. - The chat no longer runs off both edges of the screen at an accessibility text size. With the text size set to Accessibility XL, an approval card — and therefore the whole conversation around it, including sub-agent cards — was 502 points wide on a 402 point screen: the transcript centred the overflow, so about 50 points were cut off each side and a person read “elper” for Helper, “ssues” for issues and “ttp” for the
httptool. Two decorations caused it: the method badge (POST) and the consequence pill (sends data out) were markedfixedSize, which is right at normal sizes — it keepsPOSTfrom wrapping — and becomes a demand for ~250 points of width when the type is three times bigger. They wrap now, but only at accessibility sizes; at every normal size the card is pixel-identical. A new UI test puts the whole fixture under Accessibility XL and fails with the name and coordinates of anything sticking out: it reports six elements atx -50…452on the old code and nothing on the new one. Thehttpcomposer's Stop button was over the right edge by 38 points for the same reason, and came back inside with it. - Settings › Tools stopped showing one tool's instructions to the model. Each row's line of copy is hand-written for the person reading it, with the tool's own
description— the text written for the model — as a fallback for anything missing. Eleven tools, eleven hand-written lines: it looked complete, anduse_agentwas not among them. So the row for it printed the prompt: “Delegate a task to a sub-agent: a fresh agent with its own system_prompt, one input and ONLY the tools you list … you get {pending:true, id} and are WOKEN with the result” — 694 characters, braces and all, 324 points tall on a 16 Pro and 356 on an SE, which is over half that phone. It now says what the person gets: “Lets the model give a side task to a helper that works on its own with only the tools the model picks for it, and reports back when it is done. Up to 3 at a time.” The guard that should have caught it checked a list of six tool names somebody typed by hand, so every tool added since was invisible to it; it now asks the tools themselves, caps a row at 180 characters, and rejects braces,key:valueand SHOUTING — prompt voice, in other words. Checked against the old code first: it fails there. - The “Asks before …” note under a tool reads as a sentence again. The tags under a row — on by default, mounted for this conversation, Asks before POST, PUT, PATCH or DELETE — reads go through. — shared one row, so SwiftUI gave the long one whatever width the short one left it: a ragged three-line column about 150 points wide, floating in the middle of the row, on the 16 Pro as much as on the SE. The approval note gets its own full-width line now, under the tags, with the same accessibility identifiers.
- Deleting the model that is answering stops the answer first. Models offers Delete by swipe and from the context menu on any downloaded model, including the one talking to you at that moment — and nothing in that path waited for it.
didDeletefired an unawaited task and returned, so the files were unlinked while the model still held them, andunload()set the app to no model running while the agent kept streaming from its own reference: the screen described a state it was not in, the next turn had nothing to answer with, and Stop — which reaches for that agent — could no longer cancel the run and fell back to its timer. The order is now stop the answer · wait for the turn to settle · release the weights · delete the files, with the same wait on Forget drive in Settings › Storage for a model living on it. The confirmation dialog says what is about to happen — "loopl 2B is answering right now. Deleting it stops that answer." — and says nothing extra for any other model, because a dialog that explains the ordinary case teaches you to skip it. It is not refused: freeing space on your own phone is a reasonable thing to want.ModelDeleteUITests3, and the behaviour test was checked against the old code first: it fails there, which is the only way to know a guard guards anything. - The approval card says what it will do to you, not how risky a policy thinks it is. Every card that asks before a tool runs carried a pill reading
low·medium·high—ApprovalRequest.Risk, which is the SDK's grade for which calls need asking, printed straight at the person holding the phone. It answers a question nobody asked. The pill now names the consequence, from the two facts that decide it: whether a network host is involved and whether the thing can be taken back —sends data outfor a write to a host,makes a changefor a write that stays on this phone,hard to undofor aDELETE,only readsfor a read. The colour still carries the grade, the sentence under it still says exactly what the tool will do, and VoiceOver now hears the consequence on every card ("It sends data off your phone.") rather than only on the severe one. - A card that sends nothing stopped offering to show "what is sent". The disclosure under an approval card decided its wording from whether the request had an HTTP method, so the clipboard card —
COPY, no URL, nothing leaving the device — invited you to "Show what is sent", flatly contradicting its own card. It asks by host now, like the pill does, so a clipboard write, a reminder and an instructions change all say "Show the change" and only a real request says sent. A test pins the two to each other, so the pill and the disclosure can never disagree about whether data leaves the phone again. Measured on an iPhone SE and a 16 Pro:ApprovalTests17, the hostname still never shortened. - The Apple Intelligence card no longer hides under the Download bar on a 16 Pro. Onboarding's pick page has a fixed amount to show — three models, the network line and the Apple Intelligence card — so the only adjustable thing is the loopl mark above them. #151 rescued a small phone by dropping that mark entirely, which left a band nobody had measured: on an iPhone 16 Pro the fourth card ended 14 points below the top of the Download bar. The mark is now sized by the room the device actually has (nothing on an SE, smaller up to 900 pt, full size on a Pro Max and on iPad). It stayed invisible in both directions — the original fix was checked on an SE, and the machine that runs the tests picks the tallest iPhone it has.
- A failed download says which model it was. The downloader's sentence describes the transfer and never names what it was fetching — right inside the Models row, which names the model one line above, and useless on the chat screen, where the same words stood alone above "No model is running yet.": "Hugging Face is limiting requests right now … Your progress is kept." Whose progress, of what? It now reads "Qwen3-0.6B didn't finish downloading." first, then the reason. Paused and interrupted downloads name the model too.
- The Retry button is big enough to hit. Measured on an iPhone 16 Pro: Retry beside a failed download was 32 × 14 pt and the chat banner's Try again 58 × 16 pt, against Apple's 44 × 44 minimum — the one control you need after a rate limit, inside 14 points of screen.
ResumeandSign inwere built the same way. All four now have a 44 pt touch area (measured: 78 × 45) and the text has not moved a point, because the hit rectangle overflows the layout instead of changing it. swift testruns again, and CI can fail. Two blind spots, one cause.Tests/LooplRuntimesTests/GGUFModelTests.swiftcalledGGUFModel.renderPrompt,applyKwargsandUTF8Accumulatorunconditionally, but all three lived inside#if canImport(llama)— so on a checkout without the vendoredllama.xcframework(every contributor, and CI) the test bundle did not compile, taking all 637 tests down with it. The two pure-Swift pieces moved out of the guard intoSources/LooplRuntimes/GGUFPrompt.swift, so they now get tested everywhere; only the Jinja-engine tests stay behind it. Nobody saw this because the macOS job ranswift test 2>&1 | tail -n 120withoutset -o pipefail: the step reportedtail's exit status, i.e. always success. Fixed, with a step that prints which tests failed instead of the last 120 lines.swift test: 637 tests, 0 failures.- The test catalog cannot drift from the app's catalog in silence.
Tests/LooplTests/Fixtures/models.jsonis a verbatim copy ofApp/Resources/models.json(SwiftPM has no app bundle); the loopl-vl 2B and 4B rows reached the app and never reached the copy, soswift testcounted 6 loopl rows against the app's 8 and the hosted job — which reads the real resource — stayed green. Fixture synced, andCatalogFixtureParityTestsnow reads both files out of the checkout and names the rows that differ, with the one command that fixes it. - UI tests run after every merge. A new
uiworkflow runs thelooplUITeststarget on pushes tomainonly, so pull-request CI stays at ~12 minutes while a geometry guard that goes red is seen within the hour. It had to be:InsetsUITests.testSettingsDeleteHistoryButtonIsHittablewas red onmainfrom #120 until #170 repaired it, because nothing ran that target. The live suites in it skip themselves without a Hugging Face token, so no run launches a billed job. - Shorter words where they land in red. The new download-failure sentences are drawn in red, unwrapped, beside a Retry button in the Models list — so the reassuring half was costing a third line of alarm. “What you have downloaded is kept, so it carries on from there.” is now “Your progress is kept.”: 103 characters instead of 140, two lines instead of three on a 375 pt screen, same promise. The test pins the length, because the sentence is only as good as the row it fits in.
- A helper's row no longer claims the wrong time.
use_agenthands back its envelope the moment the child starts, so the transcript row read Helper · 4 ms directly above a sub-agent card that was 42 seconds into a background run — the 4 ms is the hand-off, and it reads as the helper's own time. A row whose result says the work is still going (pending) now shows no duration at all; the child's card carries its elapsed time. Read off the result, not the tool's name, so ause_agentcall that waited for its child keeps the number it earned. Tests:HandOffDurationTests3. - A failed download says what to do. Hugging Face rate-limits and outages reached you as the status code itself — “HTTP 429 for model.safetensors” under a half-finished download, “Hugging Face answered HTTP 503 for /api/models/…” in Browse — which never said whether to wait, retry or stop. One mapper now writes those sentences for the download path, the catalog and the Hub client: a 429 reads “Hugging Face is limiting requests right now — wait a minute and try again (429). What you have downloaded is kept, so it carries on from there.”, a 5xx “having trouble right now — try again in a few minutes”, a 404 names the file that went missing rather than its whole path, and 416 says the thing that fixes it instead of “range mismatch”. The code stays in the sentence, after the words. Tests:
HubStatusTests5. - Settings footers in two lines. Network: Offline mode stops all network use, downloads included. Reading a page runs on its own; sending or changing data waits for your tap (2 minutes, then it is denied). Hugging Face, for someone who has never heard of it: Hugging Face hosts models online. Sign in to download your private ones, share conversations and train there. System prompt: no tool name — Ask for a standing rule and the model can add it here — each change asks you first.
- Re-audit touch-ups. History's empty state says Your conversations appear here. (a thread is kept from its first message); the chat's "is ready" hint no longer truncates to "…is ready. Eve…" at the largest accessibility sizes on an SE; Settings › Agent's footer is two lines (Runs until it has an answer or you tap Stop. It sees only the tools that are on…).
- A reopened question that never got its answer says so. Since #166 a conversation is kept from the moment you send; if the app was closed before the model answered, reopening it now shows one line under your message — The app closed before the model answered. Hold the message and choose Retry, or send it again. Also fixes
EngineArchiveTestson main (they read the archive synchronously; since #165 the app's folder loads in the background). Test. - Docs screenshots match the app again. The sixteen captures under docs/assets/app2 were a day old and showed what rounds 3–4 removed: Qwen3-0.6B (4-bit) titles,
current_timerow names, the mic on the left, the old Models order with ctx 8,192 · MLX metadata, 51 bytes body on the approval card. Re-shot from the same fixtures on current main (iPhone 16, light + dark); alt texts updated to name the rows as the app does. - A turn you lose is a turn you keep. A conversation only reached History when the model's loop ended normally — on Stop, or when you opened another thread. A model error, a tool that fell over, iOS killing the app for memory during a video turn, or swiping loopl away mid-answer threw the turn away: your message, the partial answer and any tool rows were gone on relaunch, and a first-turn failure meant the thread had never existed. Now the message is written the moment you send it, every exit from a turn saves, and leaving the app saves what is on screen. A failure also reads as one line in plain words — "Couldn't finish this answer — out of memory" — where it used to print the same reason twice (stopped: X from the agent and error: X from the app). Tests:
EngineArchiveTests5. - History loads off the main thread. Every saved conversation — pictures included — was decoded on the main actor before the first screen could paint; the folder is now read in the background and the list fills in (a spinner while it does, never "nothing here yet"). Threads saved or deleted while the read runs are kept as you left them. Measured on the sim: 200 threads with pictures, 31 ms on the main actor → 0 ms. Tests.
- History you can search and skim. Conversations are grouped by when they were last touched — Today · Yesterday · This week · This month, then one group per month — and a search field (shown once there are six or more) filters by title and by what was said; the result header counts the matches. With eight or more conversations the History sheet on iPhone opens tall instead of the half-height peek that showed three rows. Tests.
- Three plainer lines. The onboarding model cards badge sees pictures like the Models rows (was vision); the Train screen's summary is two sentences that name neither the GPU nor file formats (Runs on Hugging Face in your account; the result is saved privately as …, ready for your iPhone. — the GPU and formats stay under More options, which is what Fine-tune manually is now called, under the header How it trains); download percentages follow the region's format (%40 on a Turkish iPhone). Tests updated.
- Tool cards at small sizes and big text. An approval card for
notes.example.comshowed notes.examp… — the risk pill squeezed the hostname, the one word you must read before tapping Approve. The host is never shortened now: the pill moves above it when they cannot share a line. And at accessibility text sizes the composer's round buttons drew their glyph from the text size while the circle stopped at 44 pt, so Stop became a black square on black in dark mode and Send lost the orange disc that tells enabled from disabled — the glyph is sized from the circle (identical at everyday text sizes). Tests:ControlGlyphTests3; captures on SE dark and 16 Pro light, default and AX-XL. Also repaired a geometry guard that had been red onmainunnoticed (UI tests do not run in CI):InsetsUITestsstill looked for Delete conversation history on the Settings list, where it has not lived since the Storage screen appeared. Two more tests were wrong rather than the product: they asserted on625.1 MB/52.4 MB, so the suite went red on any device whose region writes625,1 MB— they now build the expected text with the same formatter the app uses. - Models —
loopl-vl 4Bjoins the catalog and the collection (the second half of the owner's "add the loopl-vl model into our collection").cagataydev/loopl-vl-4b-4bitis Qwen3-VL-4B-Instruct with a LoRA r64 (merged) trained on the same loopl-train revision as loopl v1.1 (db6fa452, 2 epochs, best epoch 1: eval loss 2.80 → 1.73; vision tower frozen and kept, 315 tensors). Gates on a Mac against the stock 4-bit base, every one ≥ base and most clearly above: quiz 137 vs 106 /527 · tools 9/9 · 9/9 = base · identity 5/5 vs 4/5 · bio 8/8 = 8/8 · identity without a system prompt 12/13 vs 2/13 · image nouns 7 vs 6 (reads "LOOPL 42" exactly) · video nouns 4 = 4; through the app'sMLXModelon Metal 4/4 (describe, read 42 →calculator, short text turn, 10 s shapes clip at 29.9 tok/s, peak 3.3 GB). Row:vision: true,thinking: false, hermes dialect, not featured, 3.11 GB from the Hub's blob sizes,min_ram_gb6 (same as loopl 4B — a 16 Pro model). The public collection is complete at 10 models (loopl 0.8B/2B/4B + loopl-vl 2B/4B, 4-bit + bf16); snapshot and Models updated. Whole loopl-vl v1 run: ≈ $16 of HF compute. - History rows read well at scale. Titles come from the first message with whitespace collapsed, cut at a word with an ellipsis, and a photo-only or video-only first turn is called Photo / Video — a message that began with a line break used to give a row whose first line was blank, and a picture alone gave an empty row. Attachment folders whose conversation is gone (a delete interrupted, a crash before the first save) are removed on launch instead of counting in Storage forever. The row's model name and the Answered by line drop the build suffix (Qwen3-0.6B, not Qwen3-0.6B (4-bit)). Tests.
- A failed training job says what to do. The card showed torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 1.2 GiB. and nothing else. Now a sentence a person can act on leads — The training machine ran out of memory. Try the smaller base model, or train on fewer conversations. (time limit, refused token, disk, rate limit, network have their own) — with the raw line under it in small type and the full log one tap away. The running/finished summary line no longer names the GPU. Test.
- Three more plain words. Load errors and banners say Qwen3-0.6B not Qwen3-0.6B (4-bit) (they used the catalog name); Settings › Network's Ask before offers Every request · Sending or changing data · Never (was Writes only); the ask_user card says Waiting for your answer (was The agent waits for your answer).
- GGUF rows by their name. The chat title read "Qwen3 0.6B · Q4_K_M (llam…" with a GGUF model loaded; the build and the engine were the catalog's name.
displayNamedrops the · Q4_K_M (llama.cpp) tail everywhere; the Models row says 396.7 MB · text only · GGUF so a GGUF twin stays apart from its MLX sibling. Test. - Wide tables say they scroll. A markdown table wider than the screen showed cut columns ("Best", "iPho…") with no hint on an SE: the edge fade only switched on after the first scroll, because it listened for scroll changes. Content and container widths are measured directly, the fade is wider and carries a small › ; VoiceOver adds "scrolls sideways".
http: a body with no method is a POST. Qwen3 0.6B asked to post a note sent{"url", "body"}and nomethod— the tool sent a GET with a body, the server answered 405, and because a GET needs no approval the card the user expects for a write never appeared. Now a non-empty body with no method means POST (so it asks first, like every other write); the 405 hint no longer tells the model to "use GET" when GET was the problem. Test.- Stop reads as one line. Tapping Stop before the model had written anything showed "Cancelled by user" in the model's voice with a grey cancelled under it — two lines for one tap, one of them pretending to be the model. The placeholder stays in the history (the model API needs a non-empty turn) but the transcript drops it and shows one quiet Stopped. Test.
- Private models ask for the account first. Tapping a private (gated) model in Models while signed out used to start the download, fail on the first byte and leave a red "…is private or gated" row under On this iPhone with a Retry that could never work. Now the Hugging Face sign-in sheet opens first (the same one onboarding uses) and the download starts when the token is in; a row that did fail that way says A private model — sign in to Hugging Face to download it with Sign in instead of Retry. UI test.
- Words that small models copy, and words people wait on.
current_timeno longer names an example city — measured on Qwen3 0.6B, "what time is it right now?" becamecurrent_time(timezone: "Europe/Istanbul")on a phone in New Jersey; the description now says call it with no arguments unless the user names another place. While the model reads a long prompt the pending row says Reading your message and the tools… after 4 s and Still reading — a big prompt takes longer on this device. after 20 s (was "…" for the whole wait). loopl VL 2B gets a plain blurb (Watches video and sees pictures — the 2B tuned for clips) and its title keeps the capitals. Settings › Look & feel has one short line under each switch instead of a six-line footer; About lists Fast engine (MLX) and Older-format engine (llama.cpp). VoiceOver: the Pause button under a download says Pause download (it used to read the percentage). Tests:PlainWordsTests3. - Models — one number, plain ears. A loopl model shows the same size everywhere (1.75 GB in onboarding, On this iPhone and loopl models — the collection row used to show the weights file alone, 1.72 GB). VoiceOver reads a Models row as "loopl 2B, 1.75 GB, sees pictures, loaded" — context length and engine moved to the detail sheet in build 17 but were still spoken. Rows under On this iPhone no longer carry a ✓ (the section says it; two ticks read as two selections) — only the ⚡ for the model in use. Collection badge sees pictures (was vision); fit badges say fits in memory / tight on memory / too big for this device's memory; Train your own reads Teach a model from your conversations. The chat's empty state with no model running has a Pick a model button that opens Models (was bold text). Fixture
--design-fixture nomodel; 2 UI tests. - Onboarding, first real run. The Use Apple Intelligence card is on screen without scrolling on a 16 Pro-class phone and peeks above the Download bar on an SE (the hero mark steps aside there; the pick page's lock line moved to the download page, the story just said it). While the last bytes are checked the title says Getting ready · Checking the files, not Downloading with nothing moving. A model that cannot start shows ⚠︎, "loopl 2B couldn't start on this iPhone." with the reason under it, Try again and Choose another model — the Hugging Face pitch stays out of that screen. Choose a different model under a running download pauses it (Models keeps the part). VoiceOver: story headline + line read as one sentence pair; approval cards say "Asks you first: POST api.notes.app. Sends data to…" instead of "approval needed: … risk medium"; tool rows say "Look at picture, running" instead of "tool describe_image running". Fixture
--design-onboarding-fail; 2 UI tests. - Models —
loopl-vl 2Bjoins the catalog and the collection (owner: "add the loopl-vl model into our collection, post tune with the same dataset").cagataydev/loopl-vl-2b-4bitis Qwen3-VL-2B-Instruct post-tuned on the same loopl-train revision as loopl v1.1 (db6fa452, 6,177 conversations, full fine-tune, 2 epochs, vision tower frozen and kept — 315 tensors in the mlx_vlm export), so it reads pictures and video natively and speaks the loopl tool manners. Base chosen by a three-arm A/B on the Thinking question: Instruct with no think blocks won (tools 9/9 both dialects on the smoke; a Thinking base taught to close<think>at once gained nothing, synthesised one-line reasoning taught it to stop inside the think block). Gates on a Mac against the stock 4-bit base, every one ≥ base: quiz 75 vs 65 /527 · tools 8/9 vs 7/9 (native) · 7/9 = 7/9 (loopl JSON) · identity 4/5 vs 2/5 · bio 6/8 vs 5/8 · image nouns 6 = 6 · video nouns 5 = 5; through the app'sMLXModelon Metal: describesvision-test.png, reads 42 off it →calculator, answers a short text turn, andMLXVideoTestsreads the 10 s shapes clip (10 frames ≈ 560 tokens, 47.6 tok/s, peak 2.0 GB). Row:vision: true,thinking: false, hermes dialect, not featured (the owner decides), 1.80 GB from the Hub's blob sizes. The collection snapshot and Models carry the new line (4-bit + bf16, both public; no GGUF — llama.cpp cannot see); the Hub cards are brand pages with a video badge and themlx_vlm --videoone-liner. loopl-vl 4B (LoRA r64, same data) is training and lands the same way when it clears its gates. - iPad says iPad. Settings › Storage read "None on this iPhone" and "Keep models on the iPhone…" on an iPad — both use the device name now. The footer under loopl models is the app's own sentence ("Made for loopl: tools, memory and pictures, tuned to work well on a phone…") instead of the Hub collection's description ("MLX 4-bit (vision) = what the app downloads, bf16 = post-tuning base").
- Models — small things. The section over the Train your own row is now headed Your own model instead of repeating the row's title; a downloading row shows 40% in its state slot instead of a spinning ring next to the bar and the byte count (three progress affordances read as a hang).
- History, sharing, onboarding — three small readability fixes. The selection bar no longer cuts its action to "Share to Hugging F…" on a 375 pt screen (and no longer shrinks the text 20 % to fit on a 16 Pro): the button reads Share when the full label does not fit at full size (
ViewThatFitsaround the button alone — aSpacerinside a candidate makes it never fit), full label on wider screens; VoiceOver always says Share to Hugging Face. The share-review footer says "Each conversation is shared as it happened: your messages, the model's answers, and the tools it used" instead of "Rows follow loopl-train's schema…". The onboarding story says the chip in your pocket, not the GPU. - Settings, Storage, Hugging Face and approval cards in plain words. Settings › Agent: "The model keeps working until it has an answer or you tap stop. It only sees the tools that are on, and can switch one on for a single conversation when it needs it." (was the loop … mount … manage_tools); Network: "reading a page runs on its own; anything that sends or changes data waits for your tap (2 minutes, then it is denied)" (was GET and HEAD run; POST, PUT, PATCH and DELETE…); Look & feel footer explains each toggle in a sentence a person can use (how fast the model wrote, not tokens per second; GGUF builds: an older, text-only model format). Storage: drive not volume, models not weights, Diagnostics connecting, reading, writing. Hugging Face: one sentence for the token everywhere (read downloads your private models; write also shares and trains; stays in the Keychain). The http approval card says Sends data to notes.example.com · 3 headers / Reads a page from … / Deletes something at … instead of Sends a POST request · 51 bytes body, and its disclosure reads Show what is sent / Show the change instead of BODY 51 B (the byte size moved to the VoiceOver label). The glossary is now in docs: Design › Words.
HttpToolTestsasserts the new detail lines. - Train your own — written for a person, not an ML engineer. The form no longer says "The training script is committed next to the shards as train/loopl_sft.py", "A small LoRA adapter is trained and merged in", "held-out chats", "no bf16 checkpoint next to the 4-bit weights — training from 4-bit is refused" or "Guide: under 200 conversations 3–4 epochs". It says: Private, in your Hugging Face account · Starts from loopl 0.8B and learns from your 191 conversations in up to 3 passes, keeping the one that scores best. It runs on an A10G 24 GB at Hugging Face, in your account; the result is saved privately as …, ready for your iPhone · loopl-old can't be continued — it was not saved in a form that can be trained further · Passes instead of Epochs (Auto tries up to 3 passes and keeps the one that scores best. Fewer conversations need more passes…). Stages read Reading conversations · Preparing · Training · Finishing the model · Checking · Uploading · Making the iPhone version · Making the Mac and PC version · Done; the running card drops LoRA and stitched; the result card says Score by pass — lower is better and Kept pass 2 — it scored best. The method still shows under Fine-tune manually; ids and the approval body (the real command) are unchanged.
TrainingTests+1 (stage labels carry no jargon), plan-summary assertions updated. - Models — rows say what matters. Rows under On this iPhone read 1.75 GB · 👁 sees pictures or 335.5 MB · text only instead of 1.75 GB · 2B · ctx 12,288 · MLX (context and engine stay in the VoiceOver label; the size class is in the name). loopl models rows use the same one-line blurbs as onboarding ("Smallest — quick answers and tools, sees pictures.") instead of the Hub's curator note ("MLX 4-bit, vision: … 297 vision tensors. quiz 158 vs base 83 · tools 9/9"); other collection rows show the note's first sentence with build clauses removed (
ModelBlurb.plain). The chat title and ready/loading lines useModelSpec.displayName— Qwen3-0.6B instead of Qwen3-0.6B (4-bit);nameis unchanged for History and sharing. Browse row: Any public model on Hugging Face — no account needed. Tests:ModelSpecDisplayNameTests,ModelBlurbTests(app target). - Chat — the title bar is legible over a long conversation. On iOS 26 the transcript scrolled under the translucent bar with the default soft edge, so the first lines of a long chat read straight through the model name and the clock (clearly on an SE in dark). The transcript now uses the system's hard scroll-edge effect at the top (
HardTopEdge, a no-op before iOS/macOS 26): content stops at the bar, the title stays readable, nothing else about the layout changes. - Fixed — the stop button, and
locationon the phone (STOP-LOCATION; owner on build 15: "stop button is not working properly, and the location tool is not working in the ios"). Two SDK bugs with one root: a tool parked on a continuation that never resumes. Stop:Engine.stop()cancelled the Task iteratingagent.stream()— anAsyncThrowingStreamends withnilfor a cancelled iterator, so.resultnever arrived andisRunningstayed true; and a tool that ignored cancellation kept the agent's turn alive anyway. NowAgent.cancelRun()cancels the run from the host (the host keeps receiving every closing event),runToolraces the tool against cancellation → Tool execution cancelled at once (the late result is dropped),ModelGate.acquire()throws when cancelled while queued (a parent stopped behind a child's model call no longer waits for it), a model stream that ends withnilunder cancellation is reported.cancelled, andTranscript.stopped()settles open tool rows / approval cards / running sub-agent cards as the host's 1.5 s fallback.StopTests(hung tool → turn over < 1 s with a well-formed history, mid-text, gate, transcript). Location:CoreLocationSourcecreated itsCLLocationManagerlazily from a tool-executor task — a thread with no run loop, and CoreLocation delivers delegate callbacks on the creating thread's run loop — sodidChangeAuthorization/didUpdateLocationsnever fired: the permission answer and the fix went nowhere and "where am I?" hung (which the broken stop then could not end). The manager now lives on the main thread; every wait has a deadline (10 s for a fix, 60 s for the dialog) and observes cancellation; errors say what to do (allow it in Settings › Privacy & Security › Location Services › loopl, try outdoors).locationis on by default (coarse ~1 km + the iOS permission already gate it); installs that had it off get it switched on once (AppleToolPrefs.laterDefaultOn), switching it off again is respected.ResumeOnce(one continuation, resumed at most once from any thread) is public inLooplfor device tools that wait on delegate callbacks. - Controls on tool results (owner: "use the flashlight-like UI components to be able to close them early"). SDK: additive
ToolOutput.control→ToolResult.control, a typedToolControl(toggle · button · player · mapPin · timer · link, flat{"type": …}JSON, tolerant decode, never in the model's text) that the tool attaches;Agent.invokeTool(_:)runs one tool for the user through the normalrunToolpath (hooks, approval policy, mount-on-call) with no model call, and records it in the model's history as a tool pair prefixed [the user tapped this in the app];Transcript.applyControlResultupdates the tapped row in place (controlTaps). On-device tools:flashlight→ Turn off / Turn on (newTorchSourceseam,FixtureTorch),vibrate→ Vibrate again,music→ mini player (transport + volume → the tool's ownactions),location→ map pin at the model's rounding,calendar_create/reminders_create→ Open in Calendar / Reminders. LooplUIToolControlView(chips, player, MapKit pin that opens Maps, countdown, link;looplControlHandler/looplControlBusyenvironment; disabled when no model or a turn is running). App:Engine.tapControl, History saves the control, fixture--design-fixture controls(--design-controls <tool>) runs the real flashlight/music code through the seams. Docs: Tools › A result can carry a control, Controls on a result. Tests: ToolControlTests 6, ToolControlsTests 5, ToolControlsUITests 2 (tap Turn off → flashlight: off → Turn on; player Pause → Play). - Chat — plain words instead of developer names. Tool rows are titled Time, Calculator, Copy to clipboard, Tidy context, Helper, Card instead of
current_time/clipboard_write/manage_messages/use_agent/render_ui(ToolLabel.display, one table in the SDK; an unknown id reads as its words —fetch_weather→ Fetch weather); the id the model called now sits at the top of the expanded card, and History, the dataset and VoiceOver labels are unchanged. The helper wake line reads "Helper issues finished · 1 min 13 s" instead of[use_agent agent_i finished] issues · done · 73 s(the bracketed form still goes to the model); the tidy marker reads "The model tidied 2 turns of notes · your conversation is unchanged"; system-prompt notes read "The model added a rule to its instructions · undo in Settings › System prompt".ToolLabelTests(4); existing assertions updated. - Approval card — you can read what you are approving. Non-http cards used to pack the tool and the payload into one truncated line ("Add reminder · Call the…", "Copy to clipboard · Altı…"); the label is now the heading and the payload gets its own line that wraps up to three lines (
ApprovalCard.titleParts; http cards keep host + path). Calendar and Reminders cards say the date like a person — "Adds to Reminders — tomorrow at 9:00 AM", "Fri 9 Oct at 15:00", the year only beyond a week (AgendaStore.spoken) — instead of2026-10-07 09:00. Tests:ApprovalTests.testCardSplitsLabelFromPayloadForNonHttpTitles,AgendaToolsTests.testSpokenDatesReadLikeAPerson. - App — send a video. With a Qwen3-VL model loaded the composer's + offers Photo or Video (the Photos picker takes both; a movie arrives as a file, never in memory), Record Video (camera, 480p) and Video from Files; one clip per message. The chip shows the poster ·
m:ss·N frames· listening… / with transcript / no speech. The clip is copied into the thread's attachments folder (deleted with it), probed, and the frame plan is sized to this phone:MemoryBudget.maxVideoFrames(VisionBudget) caps the count,VideoFramePlan.fittingkeeps the frames inside half the context window. The audio track goes through the same on-devicetranscribeas voice notes and travels as[audio transcript: …]. A text-only model keeps the clip in the message and reads the honest stand-in + transcript, with a note in the transcript. History keeps the block (optional field — old files decode), the user turn rendersVideoClipChip+ a collapsible transcript row, Share to Hugging Face writes a reference (row.videos = [{ref, duration, width, height, frames}]) and the stand-in text — never the bytes. Fixture--design-fixture video,VideoInputUITests(2). Measured on 2B/4B × Instruct/Thinking (table in Video input): 112 tokens per frame pair, +0.2 GB for 16 frames, motion is read by the 4B (Thinking at 10 frames), the 2B names colours and positions. Nodescribe_videotool — decided by measurement. - Models — Qwen3-VL 2B / 4B Thinking join the catalog (owner: "use the VL models in loopl — Qwen3-VL-4B-Thinking / 2B-Thinking"). No quantisation work was needed:
mlx-community/Qwen3-VL-{2B,4B}-Thinking-4bitalready exist with the vision tower, and mlx-swift-lm 3.32.3 runs them (Qwen3VL.swift). Rows carry the Hub's blob sizes (1.80 GB / 3.11 GB),vision: true, the newthinking: true(ModelSpec.isThinking→ a 🧠 thinks badge on the row) and noenable_thinkingkwarg — because, read from the checkpoints'chat_template.jinja, the Thinking template has no such switch: it always ends the generation prompt with<|im_start|>assistant\n<think>\n, so the model's output starts inside the think block and never emits<think>. The stream gate used to enter reasoning only on a literal<think>— reasoning would have streamed as prose and</think>would have leaked. Runtime fix:MLXModeldecodes the rendered prompt's tail (PromptEncoder.tail/endsInsideThink) and startsToolCallStreamGate(startsInThink: true)— works for any thinking checkpoint (a future loopl-vl included) with no per-row flag. Pre-existing bug found by the new tests:ToolCallParserextracted<tool_call>before stripping closed<think>blocks, so a call the model merely drafted while reasoning was executed; closed think blocks are stripped first now. Proven on Metal through the app'sMLXModel(MLXVisionTests on all four Qwen3-VL rows): every one describesvision-test.png("a red circle and a blue square … LOOPL 42") and reads 42 off the picture →calculator("42*3"); Thinking rows land ~300 chars of reasoning in the reasoning row and nothing leaks into the text (2B Instruct 45 tok/s · 2B Thinking 44 · 4B Instruct 35 · 4B Thinking 30.5 on an M-series Mac).LOOPL_PROMPT_MANUAL=1is a measurement switch that forces the manual loopl dialect so it can be scored against the native template tools.VisionBudget(SDK, pure): what a picture or a video costs in tokens and KV bytes for the Qwen3-VL geometry — 32 × 32 px per token (patch 16 × merge 2; a 1 MP photo ≈ 981 tokens), video frames pair up (temporal patch 2: 16 frames at 448 × 336 ≈ 1,178 tokens), KV 112 KiB/token on 2B and 144 KiB on 4B,bytesNeededandmaxFrames(0 = this video cannot go to this model here);MemoryBudget.fitsVision/maxVideoFramesconsume it and P8-B's video pipeline caps frames with it. docs/models.md regenerated. - SDK —
render_uimeasured on loopl-0.8b/2b and made to survive what they emit.RenderUIMeasurement(real MLX, app prompt, 8 portable tools, 3 runs): loopl-2b chart 1/3 · map 1/3 · diagram 0/3 → after this PR 2/3 · 3/3 · 0/3; loopl-0.8b map 0/3 → 3/3. Almost every miss was the right content in the wrong wrapper, soRenderSpecnow repairs — and reports, as notes for the next call —componentssent as a JSON string (also whenToolValidationhas wrapped it into["[…]"]), an extra}or a truncated tail (bracket stack rebalance), a key missing its closing quote ("points: [), a missingtype(fromkind, from a type-named key like{"chart": {…}}, or from the fields),edges/pins/seriesbeside the component instead of inside, and a diagram split across edge-less components. The result line says the card above is already drawn — do not send it again: a line labelledissues:made the 0.8B re-send the same card four times. Tool description leads with one JSON example per headline kind and call it directly (no calculator first) — the map example alone took loopl-2b from 1/3 to 3/3. What is left is training: calculator-first on "chart …", diagrams without edges, re-listing the numbers after the card (DATASET-V12 M1–M10 in loopl-train COORDINATION).RenderUITests18. - SDK + App — the agent edits its own instructions:
system_prompt(owner: "we also need to add the system_prompt tool to update own system prompt").SystemPromptTool {action: get | set | append | reset, text?}behind aSystemPromptStore(the app's is Settings › System prompt;InMemorySystemPromptStorefor tests). The prompt is global by decision — it is the one field every model gets. Every write is HITL at medium risk and the card shows the change itself (− old / + new; append:…790 chars kept / + rule) —ApprovalCardnow opens a preview by default and labels it CHANGE when the request has no method/URL (http, clipboard and agenda cards are unchanged). Approved →Agent.changeSystemPromptmakes it the prompt of the next model call of the same turn (same request-time rule asmanage_tools) and the store persists it for every later conversation;AgentEvent.systemPromptChanged(SystemPromptChange)→ one quiet transcript note "system prompt extended by the agent · revert in Settings › System prompt". Denied →ERROR:the model reads, nothing moves; a rule already present / already the default skips the card. App: on by default (installs from build ≤ 14 get it once vialaterCoreTools), Settings › Tools row asks first;AppSettings.systemPromptOrigin(the app default · edited by you · changed by the agent · 8:25 PM) and a 5-entry history → Settings › System prompt leads with the attribution and Revert to the default prompt / your prompt / the previous prompt; the row in Settings carries the caption. Fixture--design-fixture sysprompt(--design-sysprompt approved),SystemPromptToolTests(5, incl. call n old prompt / call n + 1 new prompt),SystemPromptUITests(3). Schema posted for the v1.2 dataset lane in COORDINATION. - SDK + app —
render_ui: dynamic native UI from a typed spec (owner: "in iOS we can generate UI components dynamically using native Swift — maps, graphs, diagrams").RenderSpec(args:)parses{title?, components:[{type, …}]}with eleven kinds —markdown·text·rule·keyvalue·table·list·chart(bar | line | pie) ·map(pins + route) ·diagram(nodes + edges) ·progress·image(url | attachment) — leniently (numbers as strings,"4k", object rows,"a -> b"edges, nodes named only by edges) and records every repair inissues, which the model reads back in the result (rendered 2 components (chart, map). The user sees the card — do not repeat the same data in prose.); an unknowntypeis kept as.unknownso the card can say so.RenderUITool(render_ui);image{attachment}hands the user's picture back throughToolOutput.images. LooplUIRenderCard: the ordinaryrender_uitool row with the card under it — Swift Charts (bar with series, catmull-rom lines, donut pie with its own value legend), MapKit (Markers +MapPolyline, tap → Apple Maps with all pins), a Canvas diagram with layered layout (DiagramLayout: DFS back-edge detection so a retry loop does not invert the flow; long labels wrap, a row shrinks proportionally) and arrowed edges,MarkdownTable/MarkdownBodyfor tables and prose; every kind carries a spoken summary for VoiceOver; AX sizes stack the key/value grid. Cards re-render from the savedToolUse.input, so History shows them unchanged. On by default (laterCoreTools), Settings › Tools row;--design-fixture render --design-render chart|line|pie|map|diagram|table|keyvalue|list|progress|misc|unknown|mixed.RenderUITests(11) +DiagramLayoutTests. → Render UI (SDK) - SDK + app —
ask_user: the agent asks you with a native form (a port of tiny-tech'suse_interact).AskUserTool {type: confirm|select|multiselect|input|buttons|form|message, text, title?, options?, default?, fields?: [{name,label,type: text|password|number|toggle,default,required}], timeout_s?}raises an interrupt namedask_userthroughInterruptScope.interrupt— the very mechanism the approval card uses, no fork: the agent emits.interruptRaised, parks that one call, and the tool runs again with the answer onceagent.resolve(interrupt:)arrives. Result JSON{ok:true, value}(bool · option ·[options]· text ·{name: value}·"acknowledged"),{ok:false, cancelled:true}on Cancel,{ok:false, timeout:true}aftertimeout_s(default 180 s, clamped 5…3600) — new optionalInterrupt.timeoutcarries a per-interrupt wait (InterruptContext/Scope.interrupt(…, timeout:)), the agent falls back toapprovalTimeoutwhen nil. Loose specs are normalised (string options,{value,label},["v","Label"]; a field called password is a password); bad specs areERROR:results the model can repair.TranscriptgrowsKind.ask(toolUseId:, AskUserRequest, answer:): the tool's own row becomes the card when the interrupt is raised, settles on.interruptResolved, and the tool result folds in (no duplicate row);transcript.pendingAsks. Privacy rule:AskUserAnswer.masked(for:)replaces password values with••••before anything but the tool result sees them — the card, History (saved as theask_usertool pair, reopened as the settled card), a rebuilt history and the HF export all carry the masked answer; theSecureFieldopts out of iOS's Save Password? sheet.LooplUI.AskUserCard: Yes/No · radio list + Choose · checkboxes + All/None + Done · text field + Send · chip row (FlowLayout) · grouped form (text / password / number pad / toggle, required check) + Submit · note + Continue; waiting for you pill, Cancel always, settled = one quiet line with the question a chevron away; AX sizes stack; ⌘↩ sends/submits;AgentCardState.waitingnow also covers a pending ask inside a sub-agent (answered on the child viaSubAgentHub.resolve(child:interrupt:response:)). App: on by default incl. old installs (laterCoreTools),Engine.answer(ask:), fixture--design-fixture ask --design-ask <type> [--design-ask-state answered|cancelled|timeout].AskUserTests(14, incl. the real 5 s timeout through anAgent),AskUserUITests(8 answer flows). Docs: Asking the user. - SDK — video input (
VideoBlock) (owner: "we will send videos to the agent and it will understand the video natively").ContentBlock.video(VideoBlock): the clip stays a file in the thread's attachments folder; the block carries duration, size, a poster thumbnail, aVideoFramePlan(maxFrames·fps ≤ 2·maxEdge448) and the transcript of its audio track — Qwen3-VL has no audio input, so the audio goes through on-devicetranscribeand reaches the model as[audio transcript: …], visibly.MLXModel's vision path now fires on images or videos and feedsUserInput.Video.framessampled by loopl's ownVideoFrameSampler(AVAssetImageGenerator, evenly spaced, ≤maxEdge) — mlx-swift-lm 3.32.3's sampler has an unbounded frame cap and one pixel budget shared with photos, so loopl bounds both.VideoFramePlan.estimatedTokens(one token per 32 px block, two frames per temporal patch) andfitting(tokenBudget:)size the prompt before decoding;ContextWindowOverflowErrornames the frame count and suggests fewer. Text-only models readMessage.textWithVideoFallback(the frames need a vision model; this model cannot see them + the transcript).TokenEstimator/HistoryEditorknow the block. Docs: Video input. Tests:VideoBlockTests(7),VideoFrameSamplerTests(4, synthetic fixture clip), liveMLXVideoTests(MEASURE line: tokens · tok/s · peak memory). - Default system prompt 1.1.0 (the directives live in the user-editable prompt, not in code): mount a missing tool with
manage_toolsbefore saying you can't ·manage_messagesstats/compact when the conversation gets long ·use_agentfor side or long tasks (a background agent wakes you). Mirrored verbatim in loopl-traindata/system/app-default.md1.1.0 so the v1.2 rows train under the prompt the app really sends. An unedited stored 1.0.1 default upgrades on launch, exactly as the build-9 one did (#97); edited prompts are kept. No "write tool code" directive — runtime tool creation was skipped by owner decision. - SDK —
ToolRegistry+manage_tools: the agent manages its own tools, effective on the next model call (owner: "when we use the tool named manage_tools, that will be immediately effective for the language model call we will make").ToolRegistry(actor) = the catalog of every known tool — mounted or not, runnable here or not with the reason — plus the mounted set in catalog order;mount/unmount/register/unregister/setMounted/list, aversionand observers.Agent(…, registry:)makestoolsa mirror ofregistry.mounted, re-read before every model request and before every tool execution (syncTools) — so a tool mounted in cycle n is offered AND runnable in cycle n + 1 of the same turn;AgentEvent.toolsChanged(mounted:)announces each change;Transcript.toolsMountedkeeps the last one.ManageToolsTool(list | mount | unmount | describe— nocreate: by owner decision loopl never loads model-written code).ToolRegistryTests(10) include the proof: three requests on one user message, request 2 carrieshttp,httpran. No registry → the agent is byte-for-byte what it was. App: one registry per Engine — Settings › Tools is the initial mounted set of every new conversation,manage_toolsis pinned (always on), the thread's mounted set is saved with it (SavedConversation.mounted) and restored on reopen, Settings › Tools tags rows mounted / unmounted for this conversation and a line says what the agent changed; a Settings switch applies to the current conversation too. Measured (ManageToolsMeasurement, needed tool unmounted, 3 runs): v1 plain description 0/12 on both loopl-0.8b and -2b — the model calls the missing tool by name, reads Unknown tool, stops; so the unknown-tool result for a known-but-unmounted tool now says the exact mount call → loopl-2b mounted 11/12, used it the same turn 10/12 (http 1/3 only because of the harness's 5-cycle cap afterweb_extractorguesses; time 2/3, calculator 3/3, remember 3/3). loopl-0.8b reads the hint and then tells the user instead of acting (time 2/3, rest 0/3) → the app setsToolRegistry.mountOnCall: a direct call to a known, unmounted tool mounts it for the conversation and runs it, result prefixed (X was not mounted — mounted it for this conversation.). - LooplUI / sub-agent cards: every
TranscriptItem.Kind.agentrow (one per child, right under itsuse_agenttool row) renders as anAgentCard— name, one line of system prompt, the tools it may use, a status chip (AgentCardState: queued · running · waiting for you · background · done · failed · stopped — waiting for you is derived from an unanswered approval in the child's transcript), elapsed (ticks once a second while it runs, frozen after), and the child's own transcript rendered with the parent's components (StreamingTextView,ToolRow,ApprovalCard,ReasoningRow) — a child's approval card answers through the sameonApprovalhandler (the Engine bubbles it to the child). Inline children are open while they run and fold to their result line when over; background children fold to a chip line (tap to open) andBackgroundAgentsStripkeeps that line under a settled transcript (tap scrolls to the card). Fixture--design-fixture agents(--design-agents finished|live),AgentCardUITests(2),AgentCardStateTests(4). Replaces the SDK PR's placeholder row. - SDK — sub-agents (
use_agent).SubAgentHub+UseAgentTool {action: run|status|result|stop, system_prompt, input, tools, wait, id}: a childAgenton the same loaded model with only the listed tools (validated;use_agentrefused → depth 1), events nested in the parent's stream as.childStarted/.child(id, event)/.childStatus,Transcriptgrows.agent(SubAgentRun)rows with a nested transcript.ModelGateserialises model calls across parent and children (FIFO, one call at a time, 3 children max, the rest queue).wait:falseor >45 s →{pending:true, id}; a background child finishing wakes the parent with a synthetic[use_agent <id> finished] …turn (queued while generating or typing). Child approvals bubble into the parent's UI and are answered on the child; a thread switch stops every child. App:use_agenton by default, placeholder card untilAgentCardlands. Docs: Sub-agents. - App —
manage_messageson by default: the portable tool set is 7; Settings › Tools lists it under On with on by default and no approval note (it edits only the model's working context). Installs from build ≤ 12 get it switched on once (AppSettings.laterCoreTools/loopl.toolsSeen) — a later off stays off. The transcript keeps every row and shows one quiet marker when the model compacts (.note,transcript-note); fixture--design-fixture compacted,ManageMessagesUITests2/2 (marker + rows kept + Settings row). - SDK — the agent manages its own context:
manage_messages(a port of tiny-tech's manage-messages.ts).ManageMessagesToolgives the modelcompact(old tool calls and results out, every message's text stays — default keeps the last 3 turns,keep_lastto change),drop(whole old turns by number),drop_tools(by id or tool name),clear(only the current turn stays) and the read-onlylist/list_tools/stats.HistoryEditoris the pure, unit-tested half: the active turn is never edited, an orphaned call gets a synthetic result (merged into a real carrier when one follows), reasoning stays whole or leaves with its tool blocks, and the structure is validated before the history is replaced — a bad request comes back as anERROR:the model reads, nothing moves. Effective for the next model call of the same turn: tools reach the live agent through the newToolContexttask-local (ToolContext.current?.agent, set by the agent around everytool.call), andAgent.editHistoryis the one writer (it also drops the usage baseline so the next projection re-estimates). The agent edits its working context; the user's record is untouched by construction (Transcriptis built from events) and gets one quiet marker from the newAgentEvent.historyEdited(HistoryEdit)— "2 turns compacted by the agent · your history is unchanged".HistoryEditorTests(16) incl. an end-to-end check that the model's very next prompt lacks the old tool blocks. - App — Models › Storage is back, as a block you can read (owner: "we had how much space we have in the models page, which disappeared" — #112 had folded it into a one-line caption under On this iPhone). A last Storage section: On this iPhone
2.1 GB used · 54 GB free, one row per drive (used · freeor Not connected), Memory for a model, the catalog line in the footer and a row into Settings › Storage. Free space is the real free bytes of the volume (StorageLocations.internalFree/ each volume'sfree), so the number a download needs is on the page before the download. The owner's section order above it is untouched.StorageReadout(pure wording, unit-tested),ModelsOrderUITestsreads the sixth header and the row. - App — Models tab in the order a first-time user needs it (owner: "1 downloaded models, 2 Browse Hugging Face, 3 Train your own, 4 loopl models, 5 Mine — less confusion"). Five sections, top to bottom: On this iPhone (downloaded models and Apple Intelligence, the one in use marked, a download in flight with its ring, a model on a drive with the drive's name, swipe to delete; storage facts in the footer; a calm Get a model ↓ line when nothing is here) → Browse Hugging Face (+ repos added but not downloaded) → Train your own → loopl models (the collection; the GGUF disclosure and the curated More models row fold under it) → Mine. No repo twice (dedupe by repo id). One state glyph per row —
ModelPickerRow.downloadingdraws the ring instead of the arrow while weights arrive. Fixed the row's facts wrapped on an iPhone SE (ctx / 12,288 on two lines): oneTextthat shrinks before it wraps. Fixtures--design-models shelf|empty(ModelStore.overrideState),--storage-fixture none;ModelsOrderUITestsreads the headers top to bottom. - App / Composer — pixel pass (UX-POLISH audit on 16 Pro / Pro Max / SE / iPad 13, light·dark·AX-XL·RTL, keyboard 0–5 lines): the tab bar steps out while the composer is focused — the iOS 26 keyboard is translucent and Chat/Models/Settings ghosted through its bottom row; the one-line field box is 40 pt like the round buttons (was 42 → caret 1 pt off their centre); at accessibility sizes attach + mic fold into one
+menu (Photo Library / Take Photo / Paste Image / Record Audio) and the buttons stop at 44 pt instead of 56, so an SE at AX-XL shows ~4 words per line instead of 2; the round buttons useLooplDesign.controlFill(secondarySystemFill) —secondarySystemBackgroundwas the bar's own tint in light mode and the circles were invisible. Tokens:LooplDesign.controlSide/controlSideMax/controlFill. Lane evidence testUXAuditUITests(skipped unlessLOOPL_UX_SHOTS). - Fix: the chat froze after dismissing the keyboard with a tall draft (UX-POLISH, found while recording the keyboard transition): the transcript was a
LazyVStackunder a bottomdefaultScrollAnchor, and whenever the container changed size while content ≈ container the two oscillated — main thread 100 % inLazySubviewPlacements, every tap dead, never recovering (16 Pro + SE; a Pro Max was tall enough to escape). The transcript is a plainVStacknow (a thread is dozens of rows; the anchor-follow needs real heights anyway);UXAuditUITests.testKeyboardDismissWithTallDraftDoesNotHangLayoutguards it in CI. - History › Select → Delete N. The selection bar now has two actions: Delete N (destructive, one confirmation — "This removes their messages, pictures and recordings from this iPhone") and Share to Hugging Face; Select all / Deselect all sits in the nav bar; ⌫ with a selection asks the same question on iPad/Mac. Deleting removes the threads and their attachment folders, and a deleted open thread resets the chat to a fresh one.
- App / LooplAppleTools:
contacts_search(CNContactStore name lookup → phones/emails, read-only,NSContactsUsageDescription) andphotos_recent(newest N ≤ 4 or one day → pictures on the result for the model /describe_image; works with Selected Photos,NSPhotoLibraryUsageDescription) — both off by default, both with seams (ContactsSource/PhotosSource) and fixtures (--design-fixture contacts),ContactsPhotosToolsTests+ContactsPhotosMeasurementon real models. With this the Settings list is 17 on-device tools, 10 on by default (today's eight + the clipboard pair). - Composer: mic on the right (owner, 2026-10-06):
+· field · mic · send — the controls that finish a message sit together under the thumb; measured 12/12 insets and equal 40 pt centres on 16 Pro / SE / iPad (UXAuditUITestsgeometry). The keyboard-dismiss regression test skips the tab-bar assertion on iPad (sidebar shell). - Fix: a wide picture pushed the whole chat 8.5 pt off-centre on 375 pt screens (SE, iPad ⅓ split; UX-POLISH geometry check after #117):
ImageTilesized height-first, so a 1.6:1 photo came out 352 pt wide in a 335 pt column — the oldLazyVStackswallowed the overrun, the plainVStackreported it and the composer's+sat 3.5 pt from the edge. Tiles now fit inside the column (width ≤ proposal, height ≤ 220). Also: lane-onlyUXAuditUITests.testDumpAccessibility(VoiceOver label dump) andLOOPL_UX_ORIENT; the dismiss regression test skips the tab-bar assertion on iPad. - Accessibility sizes, second pass (UX-POLISH, SE at AX-XL): History's selection bar stacks Delete over Share instead of cutting both to "Del… / Sha…"; a model row's facts line wraps at accessibility sizes instead of losing its tail ("ctx 12,2…"); Train › Conversations becomes inline rows there (a menu label squeezed a repo name to "loopl-a…ations" however wide it was); Models › More uses the shared
DisclosureChevron(RTL). - App — Train your own, simplified (owner: "Off / By day / By topic makes no sense to me; 'Stop after' isn't meaningful — I don't know when we need to stop"). The Stitch picker is gone — grouping conversations into samples is the script's decision, the app never passes it;
loopl_sft.py1.2.4 makes by day the default after an A/B on the same 74 conversations (0.8B LoRA, 2 epochs: held-out loss 0.2814 stitched vs 0.2970 separate; +4 rows, +40 s), bundle re-synced. Stop after is gone — the cost guard is automatic: twice the estimate, never under 15 minutes, rounded up to the minute, sent as the job'stimeoutSeconds, and said in the one estimate sentence ≈ $0.30 · about 9 minutes · stops by itself if it runs past 18 minutes that the form and the ApprovalCard both show (TrainingEstimate.sentence,capSeconds,cap(forEstimatedSeconds:);TrainingPlan.hoursCapremoved,exceedsCapgone — the cap grows with the data instead of truncating it). What is left passes a one-glance rule: Base model · Continue a model · Epochs: Auto (recommended) · Model name, with the 1–4 picker + heuristic, GPU: Auto · A10G 24 GB and Method under a collapsed Fine-tune manually; Include all previous shards reads Learn from everything I've shared so far; the Recipe footer says the plan in plain words (TrainingPlan.summary(user:)). Stage ladder and result card unchanged.TrainUITestscover the fold, the sentence and the missing controls. - Train your own starts from the loopl models (owner decision, 2026-10-06): the Recipe's size picker reads loopl 0.8B · 2B · 4B and a fresh round sends
--base cagataydev/loopl-<size>— the public bf16 v1.1 checkpoints — instead of stock Qwen3.5, so a first training already carries loopl's tool dialect and identity (the script records it as round 2 of that lineage).TrainingBase.trainingRepo;hubIdstays the Qwen root for sizes and estimates; a tuned model from Mine still wins.TrainingTests.testFreshRoundStartsFromTheLooplCheckpoint. - App / LooplAppleTools:
music— the system music player as one tool (action: play | pause | next | previous | volume | now_playing,level0–100), on by default (owner: "so I can change the volume, next song, pause playing"), no approval card — every action is reversible from the Lock Screen.MPMusicPlayerController.systemMusicPlayerfor playback; volume through the slider inside anMPVolumeView— iOS has no public API to set the system volume, this is the one supported route and iOS shows its own volume HUD;now_playingreads the title through the Media Library (NSAppleMusicUsageDescription, asked only for that action). The parser takes what small models emit (skip→next,back→previous,status→now_playing,"65%",0.4→40).MusicSourceseam (SystemMusicSource/FixtureMusicSource,--design-fixture music); the Simulator has no player and says so.MusicToolTests(5) +MusicToolMeasurementon real models (oneactiontool measured, not assumed — evidence in the lane folder). - App — Settings › Tools is one place. Every tool the app can offer — the portable six, the on-device Apple ones, the new ones — lives under one collapsible Tools row in Settings › Agent, grouped by what the model actually gets: On and Off (a tool this device cannot run sits under Off, greyed, with the reason). Each row: toggle ·
name· one line of copy · on/off by default · Asks before … where an approval card applies;open_app's allowlist andlocation's Exact position sit under their rows. The separate Tools page and the On-device Apple tools section are gone (the Show GGUF builds toggle is not a tool and stays in Look & feel).ToolListing(pure,ToolListingTests) holds the rows, the grouping and the notes; the summary next to the row is the same arithmetic asToolCountSummary. - App / LooplAppleTools:
location— CoreLocation when-in-use, off by default, coarse unless you say otherwise: place name + coordinates rounded to two decimals (~1 km) + time zone; Settings › location › Exact position switches to the best fix iOS allows (accuracy/altitude added). The model has no precision argument — measured, Qwen3.5-0.8B setprecise: true9 of 9 times, "where am I?" included — and a parameterless tool it stopped calling (0/3), so it takespurpose: place|weather|time;weatherappends the exact open-meteo URL for the nexthttpcall,timethe local clock. Read-only (no card); permission just in time (NSLocationWhenInUseUsageDescription);LocationSourceseam (CoreLocationSource/FixtureLocationSource). Measured withLocationToolMeasurement, including the chain it exists for — location →httpopen-meteo with the coordinates. - LooplUI / disclosure + accessibility sizes (UX-POLISH audit): one
DisclosureChevronfor every expandable row (tool row, thinking row, approval body) —chevron.forwardso it mirrors in right-to-left layouts (chevron.rightpointed away from the content) and rotates towards the content on either side; the tool row's ✓/✗ badge box is a@ScaledMetric(fixed 16 pt, it overflowed into the tool name at AX-XL) and the running sweep's indent follows it; the inline chat title stops scaling past xxxLarge withminimumScaleFactor(0.8)(at AX-XL it was shoved against the History button and clipped). - App / LooplAppleTools:
calendar_events,calendar_create,reminders_due,reminders_create— EventKit, off by default (Settings › On-device Apple tools); reads list the coming days / open to-dos, creates stop at anADDapproval card with title, time and place. Permission is asked just in time (NSCalendarsFullAccessUsageDescription,NSRemindersFullAccessUsageDescription); a denial is a plain tool error the model relays.EventSourceseam:EventKitSourceon a device,FixtureEventSourcefor tests, the harness and--design-fixture agenda. Tool descriptions carry Today is yyyy-MM-dd (weekday) so a small model can resolve "tomorrow" without acurrent_timeround-trip.AgendaToolsTests(9) +AgendaToolsMeasurementon real models (evidence in the lane folder). - App / LooplAppleTools:
clipboard_read/clipboard_write— read what you copied (text or a link; a copied picture comes back as an image on the result fordescribe_image) and, through the approval card, copy text back (COPYcard quoting the text,risk: medium—.lowis auto-approved by the default policy, which the harness caught). Default on. New tools from here on declareAppleTool.defaultOn; a fresh install (and an update that first ships a tool) switches on only the default-on set, and Settings says Off by default — fewer tools help small models under the rest (AppleToolPrefs). Schema shape measured on real models before shipping:ClipboardShapeMeasurementran one tool withaction: read|writeagainstclipboard_read+clipboard_write— Qwen3.5-0.8B 0/9 vs 2/9 (the single tool turned "what did I copy?" into a write 4 times in 6), loopl-2b 9/9 both — so two tools ship, and the singleClipboardToolstays in the SDK;ClipboardToolTestscover parsing, the card and a real pasteboard round trip. - The loopl models are public (2026-10-06, owner decision).
cagataydev/loopl-{0.8b,2b,4b}-4bit(MLX 4-bit, vision — what the app downloads) andcagataydev/loopl-{0.8b,2b,4b}(bf16, the base for Continue a model) are public on Hugging Face, and the loopl collection is public and holds exactly those six. All three carry v1.1 (dataset v2.1: identity/bio family, file cards, repair rows): 4B quiz472 158 vs v1 111, tools 9/9 · 9/9, identity 5/5; 2B 91 vs 88; 0.8B 88 vs 59. Catalog: the three loopl MLX rows aregated: false(no account), notes/sizes say v1.1; the GGUF rows stay gated — those repos, the LoRA adapter and the dataset remain private.App/Resources/collection.jsonre-synced (6 items,private: false);docs/models.md+ README say public instead of gated;testBundledSnapshotInTheAppIsCurrentpins the six. - App — two labels that lied, found by the visual audit of build 10's main. Fixed Settings summarised the tool set as "Tools 6 of 6" directly above eight on-device Apple toggles that were all on:
Engine.allToolsis the six portable tools, while the agent getsallTools.filter(enabled) + AppleToolPrefs.active. It now counts both and leaves out what this hardware cannot run — 9 of 9 on a simulator, 14 of 14 on an iPhone (ToolCountSummary,Engine.toolSummary,ToolCountSummaryTests). Fixed Models › loopl models told you to sign in twice — the key-icon row and the section footer carried the same sentence 40 pt apart; the row keeps the instruction (CollectionModel.signInHint), the footer keeps the status (the collection is private right now), and a private collection with nothing cached now shows the hint instead of a warning triangle. - App: onboarding picks our models — the cards are the catalog's
featuredMLX rows, loopl 0.8B · 2B · 4B (smallest first;Onboarding.cardsreadsCatalog.featured, the Qwen trio is only a fallback for a trimmed catalog), with loopl 2B preselected and badged Recommended · fast + smart; one line of copy per card, a vision pill, the real Hub size, and a memory fit badge fromMemoryBudget.budget(4B on a 6 GB phone: tight/too big plus one line saying why and pointing at 2B). Apple Intelligence is a full-width card under the three — "already on this iPhone, no download. Private, instant, works offline. Text only, 4k context" — tappable when available, otherwise it says in one line why not (no more footnote button). A pick that is a private repo with nobody signed in no longer fails on the first byte: Download opens an inline sign-in step (the token sheet → the download starts by itself on success), Use Apple Intelligence instead, or Pick another model; public rows skip it entirely. Fixtures--design-apple-fm,--design-ram-gb N,--design-gated(--design-mine= signed in);OnboardingUITests+4; App Store shot #10 re-shot. - Loopl / App: a Hub row's size is the size you download. The listing only sends
safetensors.total, and the old estimate multiplied every parameter bybits/8 × 1.125— so Models › Mine offeredcagataydev/loopl-0.8b-4bitas 479.8 MB while the Hub's file list says 625.2 MB:mlx_vlmleaves the vision tower in bf16.HubCatalog.mlxBytes(parameters:bits:groupSize:)reads the Hub's per-dtype counts instead (a quantised parameter costsbits/8plus its share of the per-group scale and bias,group_sizefrom config.json; everything else costs its own dtype) and lands within 0.02 % of the real blob sizes on loopl 0.8B/2B/4B,Qwen3-VL-2B-Instruct-4bit,Qwen3-1.7B-4bit,Qwen3-8B-8bitandLlama-3.2-1B-Instruct-4bit— where the estimate was ‑23 %, ‑28 %, ‑16 %, ‑33 % and +6 % out.estimateMLXBytes(params:bits:)stays as the fallback for rows without a breakdown.HubSizeTestspins it on real Hub responses. - Train your own keeps going — Continue a model trains on top of one of your Mine models (bf16 twin, round badge, lineage in the card, replay of
earlier shards by default); Epochs: Auto (recommended) with the measured guide (< 200 conversations 3–4 · 200–1 000 → 2 · > 1 000 → 1–2), best epoch kept,
early stop; Stitch related conversations by day or topic into longer multi-turn samples; the result card draws held-out loss per epoch.
loopl_sft.py1.2.3. - App: a reopened conversation says who answered it. Opening a thread from History replays the visible turns into a fresh context under whichever model is loaded now, so the old answers were written by someone else — a quiet line under the title names them (
Answered by Apple Intelligence (on-device)), and narrows toEarlier answers by …once this session adds a turn.Engine.answeredBy/Engine.continuedHere; nothing is shown for a thread started here or when the model matches. Fixture--design-fixture answeredby. - Loopl / App — tools end to end (measured with real models,
ToolsE2ETests/MemoryToolsAuditTests/MemoryProbeTests): Fixeddescribe_imagenever received the attached picture —Tool.call(_:images:)andcallOutput(_:)were extension-only methods, so throughany Toolthe default always ran (now protocol requirements,ToolDispatchTests).rememberneeds onlyvalue(key optional, auto-labelled, accepts the flat shapes small models emit) and never silently overwrites a note (keeps it and hands it back unlessreplace: true— the audit caught the model writingremember{name: loopl}when asked "what is my name?");recallwith no arguments returns everything and a guessed key that misses matches loosely (colour ~ color, diacritics) or returns everything instead of hiding it.generate_voicefalls back to the best installed voice whenpersonalis not authorised. The app's default system prompt (yours to edit) gains one sentence naming remember/recall as the only memory between conversations — Qwen3.5-0.8B went from 0/16 to 8/8 memory calls (greedy), loopl-0.8b recall 0/8 → 8/8; a stored copy of the old default upgrades, edits are kept. - App — Train your own parity:
App/Resources/train/loopl_sft.pyis pinned to loopl-train (App/Resources/train/VERSION: revision + sha256;scripts/sync-train-script.sh [--check]; a test fails on drift — it had drifted 1.1.0 → 1.2.0). Now 1.2.1: the plan passes--expect-vision(every offered base sees; a text-only export fails the job), the result card shows Vision: kept ✓ (N tensors) fromdone.out.vision. - App: onboarding v2 — the first run is a story, not a sheet. The mark draws itself on a black/paper canvas, loopl settles, one line (≤ 2.5 s, tap to skip); three live pages built from the app's own components inside a phone-shaped card — an answer streams with the network off, the real
ApprovalCardwaits for your tap (or approves itself after a beat) and thehttprow follows, a photo tile and a recording chip turn intodescribe_image/transcriberows — paged byLooplUI.StoryPager(the pages follow the finger, velocity-settled, interruptible;PageDots); then the model cards from build 8, the chosen card collapsing into the progress ring; the download's dead time offers the optional Hugging Face sign-in (token sheet →HFSession.signIn); done → the chat with the composer focused. One hero mark glides intro → header → pick → ring (stage-gatedmatchedGeometryEffectslots). Full screen; iPhone + iPad; Dynamic Type to AX‑XL (chrome capped, content scrolls); Reduce Motion crossfades; haptics per step; every stage skippable. Settings › About › Show intro again replays it. Fixtures:--design-onboarding-step story|pick|download,--design-onboarding-page N;LOOPL_ONBOARDING_DEMO=1walks it at human pace for recordings. - App: hardware-keyboard shortcuts (iPad, Mac, any iPhone with a keyboard) — ⌘N new conversation, ⌘Y History (the sidebar on iPad, the sheet on iPhone;
⌘⇧His the system Home chord on iPadOS), ⌘. stop generating (plus plainEsc, since UIKit delivers⌘.as Escape), ⌘↩ send. The composer is aUITextViewand swallows⌘↩before SwiftUI sees it, so that one chord arrives through a priorityUIKeyCommandon the app delegate (App/Sources/KeyCommands.swift) and a plain Return still inserts a newline.KeyboardShortcutsUITestsdrives all four on the 16 Pro and the iPad. - Loopl / App: the built-in identity is gone —
LooplPersona,ModelSpec.personaand the loopl identity (built in) badge are removed; Settings › System prompt is the one field every model gets, as written, with nothing in front of it. loopl's own models carry their identity in their weights (loopl-train), so the prompt is yours to edit.LocalModelFiles(vision from the files) stays. - App: Models tab shortlist — the tab leads with the loopl models and Apple Intelligence (
featured: truein models.json,Catalog.featuredfor onboarding); the curated catalog folds behind one More models row; GGUF builds leave the main list until Settings › Look & feel › Show GGUF builds (downloaded ones always stay under On this device). - App: Models › loopl models — the Hugging Face loopl collection is the first section of the tab, in collection order: MLX 4‑bit rows with the vision badge from the files and the curator's note as subtitle, GGUF under a disclosure, datasets / bf16 merges / LoRA adapters / app-trained examples hidden; tap → the Browse sheet → download. Refreshed once per launch and on pull, served from the disk cache offline ("updated …" footer) and from a bundled snapshot (
App/Resources/collection.json,scripts/sync-collection.sh) on a fresh install; anonymous users get a sign-in hint while the collection is private. The curatedmodels.jsontwins of these repos are deduped out of Recommended. Loopl:LooplCollection.slug(the one place the id lives),HubCollection,HubCatalog.collection(_:)/collectionListing(_:)/rows(for:listing:)/parseCollectionSnapshot(_:). - App: share a conversation as Markdown (a
.mdfile named after it — turns under bold speakers, tool calls as one line with the result fenced, pictures and recordings counted, never inlined; device paths never leave) or as an image (the transcript column rendered withImageRenderer, 390 pt @3×, light/dark follows the app) — from the chat toolbar's Share menu and the History context menu. LooplUI:EnvironmentValues.looplStaticRenderlays tables and code cards out in place, becauseImageRendererdraws a ScrollView as nothing. - Docs: app pages for build 8 — new App › Models (first launch, catalog, Hugging Face, Mine, gated loopl v1, vision from the files, download refusals) and App › Tools (eleven tools, Ask before,
open_appallowlist, errors the model reads); Train adds the vision E2E cost line; Share and Storage refreshed. - API reference: every public type in
LooplandLooplRuntimeshas a one-line doc comment (56 added, comments only); DocC landing pages + topic groups (Sources/Loopl/Loopl.docc,Sources/LooplRuntimes/LooplRuntimes.docc), picked up bydocs/build_docc.sh. - Docs: SDK pages catch up with main — tool failures as
ERROR:results (no repeat heuristic) and trained-order tool specs,AudioAttachmentMarker→transcribe, private repos (HFAccount,HubCatalog.mine), external storage (ModelLocation,ExternalStorage.move),GenerationTrace. - Docs: loopl.dev landing v2 — the app is the hero: real iPhone 16 Pro captures from the current build (
docs/assets/app2/, @2x WebP + PNG, a 10 s muted h264/VP9 loop of a tool-calling turn in a CSS phone frame, light and dark), one CTA Get it on TestFlight, a pinned scrollytelling section whose phone screen follows six steps (models → offline → tools → approval → vision → train), the loop-flame mark drawn and pulsing likeLooplDesign.Motion;prefers-reduced-motiongets static screens per step; no new scripts or fonts. Stylesheet bumped toloopl.v9.css. - App: first-run onboarding — instead of silently downloading 335 MB, the first launch opens one sheet: three model cards (Qwen3-0.6B · Qwen3.5 0.8B · Qwen3.5 2B with size, what each can do and a vision pill from
ModelSpec.supportsVision), a live Wi-Fi / cellular / offline line (NWPathMonitor), "nothing you type leaves the device", Apple Intelligence as the no-download option when available, then the loopl mark as the progress ring (LooplProgressMark) until the model is ready. Skip = the old default bootstrap. Shown only when no model was ever used and nothing is downloaded. Fixture--design-fixture onboarding [--design-onboarding-at 0.48]. - Loopl / App: vision is decided by the files, not the architecture name —
HubCatalog.Config.hasVisionConfig(config.jsonvision_config),HubCatalog.isVision(repo:pipeline:modelType:files:)reads the repo's file list (Hit.files,expand[]=siblings; an image-processor config = a tower, a visionmodel_typewithout one = text-only export),spec(for:config:)followsconfig.jsonfirst; after a downloadModelStorere-reads the local config.json (LocalModelFiles.hasVisionTower,vision.json) so the composer's photo button and the MLX loader follow the weights on the phone.HFDownloader.wantedfetchespreprocessor_config.json/processor_config.json/video_preprocessor_config.json. Fixed: Models › Mine badged a text-onlyqwen3_5export as vision. -
App: models.json — the loopl 0.8B / 2B / 4B MLX rows are the mlx_vlm re-exports with the vision tower (
vision: true, 625 MB / 1.72 GB / 3.03 GB); GGUF rows stay text-only. -
Fixed: "Qwen3_5Tokenizer is not supported" after downloading a model exported with transformers 5 (the
loopl-*-4bitrepos). The MLX runtime now loads an unknowntokenizer_classwith swift-transformers' BPE fallback (strict: false), and when a load still fails on the tokenizer files the Engine repairs them in place —HFDownloader.repairSmallFilesre-fetches only the small files whose git blob hash differs from the Hub'soid(tokenizer, template, configs; the weights are never downloaded again) — then retries once ("Repairing" banner). TokenizerFiles(audit, error classification,gitBlobSHA1). Bundledloopl_sft.py1.0.3 normalises tokenizer files from the mlx-community base before every push and gates on what the app can load. - App: Settings › Look & feel — Appearance (System / Light / Dark) and a Haptics switch (
LooplDesign.hapticsEnabled, also gates theHapticBus); the offline airplane in the chat toolbar is a button that explains offline mode and can turn it off; the first-launch download banner has Pause ("Download paused — resume it from Models"). - LooplUI:
MarkdownBodyrenders pipe tables (MarkdownTable: header row, hairlines, horizontal scroll with an edge fade; ragged rows padded; the|---|row swallowed), block quotes (accent bar) and horizontal rules — the shapes Qwen3.5 emits that showed as raw| a | b |lines. Streaming-stable: a table is a table from its header line, and a lone|,>or--at the end of the stream is held back for one token. - App: History keeps the pictures — the images a user attached and the ones a tool made are saved with the conversation as ≤ 512 px thumbnails (
SavedConversation.Item.user(_:images:)/.tool(…images:), optional so 0.1 files still decode), so a reopened vision thread shows what was sent; "Copy transcript" notes[n images]. Share-to-Hugging-Face stays text-only. Fixed: an answer that is just a number with a period ("42.") rendered as an empty "1." list row. - LooplUI:
CodeCardhas a visible Copy button (→ "Copied" for a moment, haptic) in its header — the language, or the line count when there is none — and fades its edges where a line runs past the card, since the horizontal scroll has no indicator. - App / LooplUI: on iPad and Mac the conversation and the composer keep a reading measure (
LooplDesign.readingWidth, 680 pt, centred) instead of running 100 characters a line across the detail column; copy that named the device says iPhone / iPad / Mac (LooplDesign.deviceName), and the default system prompt says "on this device". - App / LooplUI: chat polish from the IMPROVE audit — a failed model load is said once (the readiness banner with Try again; the inline red line is for errors while a model is loaded, and it now has a dismiss ×); the composer placeholder follows the state ("Pick a model to start" after a failure, not "getting ready…");
ApprovalCardandToolRowlay out vertically at accessibility text sizes (no more "Ap-prov-e" / "calcula-tor"); the model's reasoning renders through the foldedReasoningRow("Thinking…" → "Thought for a moment") instead of a raw italic wall; downloads start with "Starting", not "Zero KB". - App / LooplUI: a breathing mark and "…" where the answer will land, from the send until the first token or tool row (
PendingRow,AgentTranscriptView.showsPending); the empty state is centred in the free space; the running sweep under a tool row leaves the instant its result lands; the composer's round buttons scale with Dynamic Type (capped at 56 pt) instead of clipping their glyphs at accessibility sizes. - App: Settings › Storage — keep models on a USB-C drive. Grant one folder (
UIDocumentPickerViewController, security-scoped bookmark re-opened each launch), see free space per volume, choose where downloads go (falls back phone ↔ drive by room), Move to drive / Move to this iPhone with a verified copy (size every file, SHA-256 up to 64 MB, source removed only after), and a drive that is away keeps its models in the catalog dimmed and undeletable; the loaded model's drive being pulled is a named error, not a crash. Storage. - SDK:
ModelLocation,StorageBookmark,StorageManifest,StorageError,ExternalStorage(freeSpace(at:),ensureSpace,makeBookmark/resolve,move(directory:to:verify:)with rollback, streamingsha256(of:)). - Fixed: after a relaunch with models on the drive,
The file "config.json.part" doesn't exist.iOS had refused the folder grant (a bookmark to a removable volume's root, or a drive that mounted after launch); the catalog read the unlistable folder as empty, auto-started a download onto the drive, andcreateFilethere failed silently. Now a connected-but-unreadable drive is its own state (“connected but loopl can't read it — re-grant the folder”, one tap, keeps every model in place), a model recorded on a drive is never auto-downloaded, the write is probed before any download,.partleftovers beside whole files are dropped instead of forcing a re-download, grants are retried for ~10 s after launch and on every return to the foreground, andstorage-diagnostics.log(Settings › Storage › Diagnostics) records every resolve/access/listing/write outcome.ExternalStorage.readiness(of:accessGranted:),ensureWritable,orphanParts(in:),StorageError.notWritable,StorageDiagnostics. - App: Settings › Hugging Face — paste a token, it is verified (
whoami-v2) before it is kept (Keychain, this device only), shown with name · token label ·read / write / jobsbadges, removable in one tap. Models › Mine lists the account's MLX/GGUF repos (private badge) and downloads them through the same details sheet; Browse resolves private repos once signed in. Anonymous use is unchanged. - SDK:
HFDownloader(token:)/HubCatalog(token:)take anHFTokenProviderconsulted per request —Authorization: Bearergoes only tohuggingface.coand is scrubbed on the CDN redirect; 401/403 surface as.signInRequired(no token) or.noAccess(token).HubCatalog.mine(author:),Hit.isPrivate.HFAccount.signIn(_:)/signOut()/providerandWhoAmI.scopesextend the sharedHFAccount(.gatedstays as a deprecated alias). - App: record audio in the composer. The mic (always shown) turns the field into a take — live level meter, elapsed time, × / ✓, swipe left to cancel, Reduce Motion respected; ✓ makes a chip (play · duration · remove) next to the picture chips. Sending adds one plain line per clip to the message,
[audio attached: <path> (<s> s) — call transcribe with this path to hear it], which the transcript renders as the same chip; the model reads it and decides to calltranscribe— nothing is transcribed behind its back. Files are AAC 16 kHz mono inconversations/<id>.attachments/, deleted with the thread. Microphone refused → a calm inline line with Open Settings, no alert. - Removed:
read_textanddetect_language(ReadTextTool,DetectLanguageTool). Fewer tools make a better small-model agent; OCR needed a picture that was rarely there, and the models detect language themselves.LooplAppleTools.allis eight tools. - LooplUI:
AudioAttachmentMarker(the marker line, parsed back byUserTurn),AudioClipChip,StaticWaveform,LiveWaveform,AudioPreview. LooplAppleTools:TranscribeTool.fileURLaccepts plain,file://(encoded or not),~and quoted paths; the old code force-unwrappedURL(string:)on a path with a space. - Site: loopl.dev redesigned for absolute minimalism — one typeface, ink on paper, one accent, hairlines; a landing that replays an agent turn; every page rewritten at about half the length. Front-matter descriptions are quoted (an unquoted
:had leaked page metadata into the body). Design. - Removed: every repeat heuristic.
RepeatBreaker(identical call in consecutive cycles refused, two refusals ended the turn) andToolCallRepeatGuard(an already-answered call vetoed in the stream gate; the Apple FM runtime re-asked once when that left the turn empty) are gone, withToolCallStreamGate(suppress:). The loop has no heuristics: a call the model issues is a call that runs, and the model decides what to do next from complete, honest tool results — including errors, which name the missing or wrong argument. Owner decision 2026-10-06. - Fixed: a failed tool call was invisible to the model. No chat template renders
ToolResult.status(Qwen3 / Qwen3.5 / Gemma / Llama printcontentonly; the<tool_response>fold and the Apple FM transcript did the same), so a small model sawmissing required parameter 'url'…as just another result and re-issued the identical call. Every prompt path now renders errors throughToolResult.promptContent— text starts withERROR:, the fold addsstatus="error"— and validation errors end withFix the arguments and call the tool again, or answer the user without it.(ToolResult.retryHint). Apple FM stays in call mode after an errored cycle instead of being told "tools have already been run for you, answer in plain text" (which forbade a corrected retry). Measured on Qwen3.5-0.8B-4bit via the realAgent+MLXModel— see PR for the before/after table. - App: hold a message. A user row → Edit · Copy · Delete from here; Edit turns the row into an inline editor (Cancel / Save & regenerate: the turn is replaced, everything after it removed, the agent asked again with tools and approvals as usual). An answer → Copy · Retry. Disabled with a hint while the agent is running; Reduce Motion crossfades. History keeps the rewritten thread with an edited mark (
SavedConversation.editedAt). - SDK:
Transcript.truncated(before:)/truncated(through:),userTurn(before:),messages(the agent history rebuilt from the visible items — tool rows pairtoolUse/toolResult1:1, notes/approvals/reasoning stay UI-only). Transcript for UIs. - LooplUI:
TranscriptActions+AgentTranscriptView.transcriptActions(_:),EditableUserTurn,UserTurnEditor,RetryableAssistantTurn.UserTurn(menu:). Fixed: an empty.contextMenu {}on a row swallowed the long-press. - Fixed (MLX): tool specs reached the chat template with alphabetically sorted keys (
descriptionbeforename,functionbeforetype) because swift-jinja sorts plain dictionaries; measured on Qwen3.5-0.8B this alone made it pickremember/recallfor http/time asks (right tool 11/24 → 24/24 with the canonical order). Specs are now rendered throughadditionalContextas ordered Jinja values, in the order the models were trained on. - Changed:
httpreports a 4xx/5xx as an error result (ERROR: HTTP 405 method not allowed for POST … — use GET …, body kept) instead of a success result with the status buried above an HTML page. Measured: Qwen3.5-0.8B re-issued a 405'd POST 7 times when it arrived as a success. - SDK:
MLXModel.setTrace(_:)→GenerationTrace(the prompt exactly as the model saw it, special tokens kept, + raw output + rejected native calls) for audits and self-tests;Tests/LooplRuntimesTests/ToolCallAuditTests.swiftruns the realAgent+MLXModelthrough 9 tool-calling scenarios whenLOOPL_AUDIT_DIRpoints at a model. - Fixed:
http"missing required parameter 'url'" with Qwen3.5 — the model writes flat calls ({"name": "http", "url": …}) under the hermes prompt and its own XML dialect (<function=http><parameter=url>…) under its chat template;ToolCallParsernow reads both (plus{"http": {…}},tool/tool_namekeys),ToolValidationmapsURL/uri/link/href/endpoint→url, places a loose value into the one missing required key, and its complaint carries a concrete example. 31 verbatim shapes captured from Qwen3.5-0.8B, Qwen3-0.6B and Apple Foundation Models are pinned inHttpCallShapesTests.HttpToolaccepts a schemeless host (loopl.dev→https://loopl.dev).JSONValue.compactno longer escapes/. - SDK:
Tool.requiresApproval(for:),ApprovalRequest/ApprovalDecision,ApprovalPolicy(always · mutating · never + trusted hosts). The agent suspends a flagged call inside the actor, emits.approvalRequested, resumes onagent.resolve(approval:); timeout (120 s) and cancellation deny; other concurrent tools keep running. Human in the loop. - Changed:
HttpGetTool(http_get) →HttpTool(http: GET/HEAD/POST/PUT/PATCH/DELETE, headers, body, timeout, capped response).http_getstays one release as a deprecated alias. - LooplUI:
ApprovalCard(Approve · Deny · Edit) rendered inline byAgentTranscriptView.onApproval;HapticEvent.approvalNeeded. - App: Settings › Network › Ask before (every request · writes only · never); the saved
http_gettoggle migrates tohttp.
0.3.0 — 2026-10-05 · docs & identity¶
- Docs compacted: 25 pages → 10 (Home · Start · SDK ×4 · Models · App · Privacy · FAQ · Changelog · API). Every old URL redirects.
- Re-imagined visual system (design tokens): warm paper & ink light, deep charcoal dark, one flame accent, self-hosted Inter Tight, 8-pt scale. No Google Fonts request, no gradients, no glass.
- Landing page rebuilt around a real captured agent run (iPhone 16 Pro, Qwen3-1.7B, 2 cycles / 4 concurrent tools).
0.2.0 — in progress (main, 2026-10-05)¶
Driven by the first hands-on review of the app on an iPhone 16 Pro. Everything below is on main;
the TestFlight build that carries it is being cut.
Changed¶
- Event loop —
Sources/Looplis the agent loop: stream →StreamAssembler→ tools via aToolExecutor(concurrent by default) → tool-result message → recurse. No iteration cap, no loop detector. The context window (SlidingWindowConversationManager, the only limiter) and cancellation bound the loop. 44 tests on a scripted model: 25-deep tool chain, parallel tools,maxTokensrecovery, retry/backoff, window trimming, hook order. - Vocabulary —
Message {role, content: [ContentBlock]},ToolUse/ToolResult,StreamEvent,StopReason,AgentEvent,EventLoopMetrics/Usage.Model.stream(_:toolSpecs:systemPrompt:)returnsStreamEvents; tool-call markup is resolved inside each runtime byToolCallStreamGate. - Max-tokens recovery — the truncated assistant message is repaired and appended, then
MaxTokensReachedErroris thrown;agent()continues from the history. - Retries —
ModelRetryStrategy(6 attempts, 4 s → 240 s) onModelThrottledError;AgentEvent.throttle. - Defaults —
GenerationParams.maxTokens = nil(until stop), context from the catalog's 8 GB budget, temperature per model card. - App — chat is a document (no bubbles, no metrics strip; Show performance details adds a quiet footer), composer with
interactive keyboard dismiss, first-launch download of
mlx-community/Qwen3-0.6B-4bitwith inline progress and auto-reload of the last model, Models tab → a tap-to-switch list, conversation history (sheet / iPad sidebar), new app icon + launch mark, App Store screenshot sets (6.9", 6.5", iPad 13").UIBackgroundModesremoved (plainURLSessiondownloads). - Package — platforms iOS 18 / macOS 15;
LooplUIbuilds on macOS.
Added¶
- Hooks —
HookRegistry/HookProviderwithBeforeInvocationEvent/AfterInvocationEvent,BeforeModelCallEvent/AfterModelCallEvent,BeforeToolsEvent/AfterToolsEvent,BeforeToolCallEvent/AfterToolCallEvent,MessageAddedEvent. - Conversation managers —
ConversationManagerprotocol,SlidingWindowConversationManager(window 40, tool-result truncation, proactive trim at 0.7 × context,pinFirst),NullConversationManager. - GGUF runtime —
GGUFModelon the official llama.cpp xcframeworkb11401(Metal,n_gpu_layers = 99, sampler chain top-p → min-p → temperature, chat template viallama_chat_apply_template, token-by-token streaming, overflow mapped toContextWindowOverflowError);scripts/fetch-llama-xcframework.shvendors the framework (+ optional simulator slice). Catalog: Qwen3 0.6B/1.7B/4B, Llama 3.2 3B, Gemma 3 1B, SmolLM2 1.7B as Q4_K_M. - Docs — loopl.dev rebuilt: brand mark, social cards, 404, instant navigation, this versioned changelog; every SDK page
written from the sources with
// snippet: compilesamples typechecked in CI (docs/check_snippets.sh), a type-name drift gate (docs/check_symbols.sh), and the DocC API reference at loopl.dev/api.
Removed¶
AgentConfig,Budget,LoopDetector,FinishReason,ModelEvent(replaced by the vocabulary above; one-commit deprecation shims inCompat.swiftforToolUse(id:name:arguments:),ToolResult(callId:…),Message.assistant(text:toolCalls:)).- Hugging Face token: no token UI, no
Keychain,HFDownloader()is anonymous — only public, non-gated repositories are listed. - The demo (scripted) runtime and the "max iterations" setting;
FakeModelnow lives inTests/only.
Roadmap¶
- Device tok/s table for GGUF vs MLX on the same models.
- Background
URLSessiondownloads (today a locked phone pauses a download). SummarizingConversationManagerwith the local model; Apple Foundation Models nativeToolbridging.- TestFlight → App Store listing (free).
0.1.0 — 2026-10-05¶
Initial public repository: SDK, runtimes, UI kit, app and site in one night.
Added¶
- SDK —
Looplcore:Agentactor,AgentConfig/Budget(default unlimited) +LoopDetector(warns, never stops),Modelprotocol (actors),Toolprotocol +ToolSchema+ClosureTool,Message/ToolCall/ToolResult,ToolCallParserfor Hermes/Qwen, Llama 3 and generic-JSON dialects (repairs truncated calls),PromptBuilder,ModelSpec/Catalog, resumable offline-awareHFDownloader,NetworkPolicy, built-in tools (CurrentTimeTool,CalculatorTool,RememberTool,RecallTool,HttpGetTool),FakeModel. - Runtimes —
MLXModel(mlx-swift-lm 3.32.3, Metal, KV cap fromcontextLength, chat-template tools,enable_thinkingkwargs),AppleFoundationModel(iOS 26, availability reasons, manual tool prompting, snapshot→delta streaming, estimated token counts),GGUFModelstub. - UI —
LooplUI:AgentTranscriptView,StreamingTextView,UserBubble,ToolCallCard,TokenMeter,ModelPickerRow,DownloadProgressRow+ design tokens (LooplDesign,LooplCard,Pill);Transcriptreducer in the core. - App — Chat · Models · Agent · Settings tabs; 15-model catalog (
models.json); download with resume; HF token in Keychain; offline mode; Demo (scripted) switch; Increased Memory Limit entitlement;scripts/build-on-device.sh. First device proof on an iPhone 16 Pro with Qwen3-1.7B-4bit. - Docs — loopl.dev on GitHub Pages, generated models table (
docs/gen_models.py), privacy policy.