Gather requirements for release after v1.12.2 #718

Activity

1. github-project-automation

   added this to Roadmapon Aug 11, 2026
2. github-project-automation

   moved this to Todo in Roadmapon Aug 11, 2026
3. mparrett commented on Aug 11, 2026

   

   mparrett

   on Aug 11, 2026

   Last edited by mparrett

   CollaboratorAuthor

   More actions

   @nooga Suggesting this one, too, depending on the cut (documentation freshening): #717
4. mparrett commented on Aug 11, 2026

   

   mparrett

   on Aug 11, 2026

   CollaboratorAuthor

   More actions

   Probe results vs v1.12.2 + proposed dispositions

   Local tip @ f3ca5f9b (includes #628) vs tag v1.12.2, same machine (darwin/arm64). Cut-relevant subset only.

   Perf (hyperfine)

   Workload	v1.12.2	tip	tip/base
   startup lg -e '(+ 1 1)' (quiet, N=50)	9.4 ± 1.6 ms	11.8 ± 1.5 ms	1.25×
   (reduce + 0 (range 1e6)) N=20	37.8 ms	41.3 ms	1.09×
   (some pred (range 1e6)) N=20	72.4 ms	100.3 ms	1.39×
   map/filter/reduce (1e5) N=20	51.2 ms	66.4 ms	1.30×

   • #686 held: reduce is near parity; residual matches vm: entering bytecode from Go still costs more than on v1.12.2 (since PR #645) #700, not a reopened reduce is 1.75x slower than v1.12.2 — bisected to #639 #656.
   • #700 shows clearly on callback-heavy shapes (some, pipeline).
   • #663 directionally confirmed (~1.25× startup). Absolute ms differ from the issue’s 5→9 (machine floor); the ratio is the signal.

   Registrar shadows

   TestNoNewShadowedHandRegistrations and the reduce fast-path tests PASS. reduce is no longer double-registered. 12 names remain in the accepted shadow set (conj, deref, get, int, name, namespace, nth, pop-binding!, push-binding!, some, str, subs). #704’s three soft guards are still open; adjacent drain work is #696 / #697.

   WASM key input (#595) × xsofy shell

   Same shipping shell; tip vs v1.12.2 host JS:

   Bundle	Sync burst of 10 'l' via _lgKey
   v1.12.2 (single-slot)	1 accepted / 9 dropped
   tip (SPSC ring, cap 8)	7–8 accepted (3 runs); burst=20 caps at 8

   Neither path busy-spins here (both return false when full). The win is acceptance. Paced LetGoHost.sendInput movement (10 keys @ 80 ms): 10/10 accepted; COI + SAB OK.

   AOT / Gloat (#628, #607)

   • --entry-frame fib demo: OK (55 / 5702887; fib(34) ~0.11 s).
   • lg-compile without the flag: packages only, no main.go (summary still reports native entry).
   • rt / e2e / entry_frame tests: PASS.
   • Gloat -E let-go-lower-vm and -E let-go-vm with LET-GO-SRC = tip: OK (fib 10/34).
   • Gloat lower-vm with LET-GO-SRC = v1.12.2: OK.
   • Gloat does not pass --entry-frame; it keeps its skip-MainChunk + override-drain template. Ownership split holds.
   • AOT lower-vm: embedded VM-fallback bundle rejected by its own runtime — unsupported capability mask 0x1 (supported: 0x0) #607 not reproduced when encoder and embedded runtime share one LET-GO-SRC. That matches the existing diagnosis: dual-knob skew in gloat (pin vs fork checkout), not a tip #628 regression.

   Proposed dispositions for this cut

   Answering the open questions on this issue:

   1. Version: lean v1.13.0 (features + #604 Go API removal). Not a narrow patch.
   2. vm: entering bytecode from Go still costs more than on v1.12.2 (since PR #645) #700 invoke cost: accept for this tag, with the numbers above in the notes. Mitigation stays on vm: drop leaveFrame's unused *Frame parameter to unpin f from memory #706 / Design: constant-space tail calls on the explicit-frame VM (closures, multi-arity, apply) #620; don’t block the cut on reclaiming the ~1.3–1.4× callback tax unless someone wants that bar raised.
   3. AOT lower-vm: embedded VM-fallback bundle rejected by its own runtime — unsupported capability mask 0x1 (supported: 0x0) #607: known limitation / gloat-side, not a let-go tip blocker when sources agree. Link the dual-knob notes; don’t hold the tag on it.
   4. Frontier (~20 open PRs): freeze lightly — land gate/hygiene if ready (#681, #706 if cheap); keep IR/codegen and bundle-format stacks (#649, #675, #501/#624, …) for after unless already approved.
   5. Performance measurement status: ratchet is not consulted; startup regressed 5ms -> 9ms unnoticed #663 startup (~1.25×): accept with an explicit note; recovery remains under Performance measurement status: ratchet is not consulted; startup regressed 5ms -> 9ms unnoticed #663 / measurement discipline (Three gaps that let the perf gate report green through a regression #705).
   6. Three generated-primitive guards pass when their premise breaks #704 / Retire the twelve hand registrations the generated registrar shadows #696 / Drain the remaining hand registrations in pkg/rt onto //lg:native #697: not cut-blocking; drain/guard work before more //lg:native migration, not before the tag.
   7. Gather requirements for release v1.12.0 #545: still open for the already-shipped v1.12.0 cut — close or retarget so this issue is the live home.

   Upgrade-notes checklist from the probes: #604 embed API removal, toolchain go1.26.5, #595 ring contract, #628 RunExecUnit / --entry-frame split, plus the accepted #700 / #663 costs if we take (2) and (5).

   Happy to adjust any of these stances, especially (2) if #706 is close enough to fold in.
5. mparrett commented on Sep 6, 2026

   

   mparrett

   on Sep 6, 2026

   CollaboratorAuthor

   More actions

   Notes update, 2026-09-06

   Updating my own snapshot above against main @ ee55803e (2026-09-06). Enough has merged since I wrote it that several things I said are no longer true, and a few sections describe work that did not exist yet. Recording the current state here rather than editing the body, so the history stays readable.

   Counts below have a short shelf life: eight PRs merged on 2026-09-06 alone, so read every number as of that SHA.

   Range

   v1.12.2..main is now 105 squash-merged PRs across 108 commits, 73,075 insertions / 9,429 deletions over 377 files, up from 82 PRs / 319 files at the 2026-09-02 snapshot. Twenty-three PRs merged in the four days since:

   #762 #675 #779 #734 #781 #794 #773 #792 #771 #624 #502 #745 #501 #730 #723 #727 #692 #758 #760 #789 #707 #780 #756

   Five of the six PRs listed as "still open, all now carrying an approval" have merged: #692, #501, #502, #624, and #675. Only #683 remains open, now with changes requested. Twenty PRs are open, four carrying an approval.

   The arcs above are still worth leaving until the wave settles, since rewriting them mid-drain buys a second rewrite. The corrections below are a different matter: they describe work that has already merged, so they do not go stale when the next PR does.

   Two things I got wrong about scoping

   My disposition (4) has not held up. It said "freeze lightly ... keep IR/codegen and bundle-format stacks (#649, #675, #501/#624, ...) for after unless already approved." I wrote that expecting the named stacks to sit out the cut. They were all approved, so every one of them merged, and the bundle-format work is now inside the range with the format version bump attached. The outcome looks fine to me; I just want to note that the "unless already approved" clause meant the disposition never had any effect, and it is still running. The range grows on its own while the timing stays open.

   My August perf numbers went stale twice in four weeks, which is a problem with how I measured rather than with the code. Anchoring a "% vs v1.12.2" table to a moving tip does not survive a merge wave. Whenever the size and perf audit on the checklist happens, I would pin a SHA first and measure against that.

   For whatever it is worth on timing: the interop set is already partly in (#773 merged, with #776, #778, and #787 approved and open), so a later tag would make it a headline arc with its own behavior-change section, and would likely pull the divergence PRs in behind it. An earlier one keeps the notes close to what is written here. I do not have a preference strong enough to argue for.

   Bundle format: version 3 and split debug info

   A .lgb arc merged after the snapshot and appears nowhere in the behavior-change section. It is the most consequential part of the cut for anyone who ships or embeds a bundle.

   • #501: opt-in DEFLATE, and a format version bump. Compressed bundles are format version 3; uncompressed stay version 2 with the pre-change byte encoding. Magic, version, flags, and capability payload stay plaintext so compatibility and opcode-set checks run before decompression. Old decoders reject a version-3 bundle as an unsupported version, and version 2 now rejects the compression flag at the header. lg -c -z and lg -b -z opt in; declared inflated size is capped at 256 MiB.
   • #502: lgbgen --compress, default off. The committed core stays uncompressed. The measurement is the reason: 297,359 B to 89,776 B (3.31x) on the core bundle and 214,656 B off the lg binary, against +1.17 ms decode and about +1.9 ms warm boot. Turning it on is a policy call, tracked under #503.
   • #624: lg -strip splits debug info rather than discarding it. Source maps and local-variable tables move to a digest-bound .lgb.debug companion; FlagDebugSplit marks the artifact, and a companion whose SHA-256 does not match the stripped payload is rejected instead of producing misleading tracebacks. lg, standalone bundles, and lg-runtime load a co-located companion automatically; LG_DEBUG_FILE overrides the path.
   • #745: the encoder no longer aliases chunks. Function constants were bound to serialized chunks by scanning for the first chunk with identical bytecode, so two structurally identical functions with different source metadata could swap chunks. The builder's live pointer-to-index mapping is now authoritative, including on the compressed path.

   The line I would want in the upgrade notes is the version-3 rejection rule: producing a compressed bundle with tip and reading it with any older runtime fails at the header. That is the intended design, and it is also the failure that reads as bundle corruption when nobody has written it down.

   #631 (strip debug from the embedded core and -w bundles) is now in flight as PR #800. Neither is referenced above, and #800 decides whether -strip reaches the -w path in this cut.

   My perf dispositions have inverted

   Disposition (2) accepts #700's invoke cost and (5) accepts the #663 startup regression with a note. Both rest on the 2026-08-11 probe table, which no longer describes tip.

   • #700's own analysis concluded the cost is an artifact of how transitions are built rather than inherent to the explicit frame chain: in-VM bytecode-to-bytecode invocation never regressed (1.00x), and the roughly 24 ns is paid entering the VM from Go, which native seq functions pay once per element.
   • #727 then adopted PreparedCall in reduce's four fold loops and measured 50.5 ms against 58.7 ms on v1.12.2, below the release baseline rather than only recovered toward it. #723 (cold error arms out of the dispatch loop, runLoopInner text down 21%) and #730 (removing the runLoop hop) merged alongside it.
   • #706 is listed as closed. Its content merged as #730; worth saying so, or the closure reads as work abandoned.

   I would now state it as "recovered on the shapes we measured" rather than "accept the tax". What stops me restating it cleanly is a second reversal. #780 merged the bench-ratchet infrastructure (deterministic forced rebaseline, regenerated Go 1.26.5 baseline, pre-push gate), which is progress on #663 and #705, both of which the table still shows as untouched. Then #791 (2026-09-04) reported IRCompile and InitFromLGB regressed against that new baseline, attributing it to #727 allocations and #723/#730 timing.

   So I would rather the perf section say the measurement is in flight than carry my August numbers, until #791 resolves and there is a pinned SHA to re-run the battery against.

   Entry-frame defects filed after the freeze

   The known-open-items table predates four issues on the #628 / --entry-frame feature the notes plan to advertise:

   Issue	Why it touches the cut
   #796	Entry-frame binaries run -main twice — main-chunk replay, then the frame
   #783	Lambda-lifted var (__lifted0) resolves to nil and panics in ExecContext.deref
   #782	Native lowering: closures in loops capture the mutated slot, not the per-iteration value
   #721	IR CSE merges var reads across set!; :mutated-vars only ever contains nil

   #796 is a correctness bug in the binaries --entry-frame produces, which is the one I would want closed before the notes advertise the feature. #724 frames what an AOT front door still needs, for context.

   Behavior changes I had not recorded

   Three consumer-visible semantic changes merged after the snapshot:

   • #758: symbols are invokable for collection lookup, as keywords already are (resolves #757).
   • #760: multi-arity dispatch routes an exactly-min-arg call to the variadic variant, and an empty kwargs rest stays nil through the map destructure ({:as o} binds nil, not {}), matching Clojure. Resolves #759.
   • #756: the var special form reaches private vars across namespaces, matching Clojure, where privacy is enforced only on bare qualified symbol references. Bare-symbol resolution stays privacy-enforcing on both the bytecode and IR paths.

   #789 also gates untagged builds against the GL and Ebitengine subsystems via make check-default-deps and a CI job, motivated by #788.

   What I got wrong in the WASM and TinyGo arc

   The arc reads "TinyGo lane (#522) with CI follow-ups (#682, #685)". Two fixes.

   #682 was never a PR. It is an issue, closed as completed on 2026-08-25. #685 is the CI follow-up that merged.

   The arc is wider than one line. In range: #522 (the lane itself, merged 2026-07-23), #685, #629 (SameFile stub for js/wasm), #595 (SPSC ring for SAB key input), #572 (RGBA surface capability), #743 (wake Plan 9 key reads on terminal resize), and #792 (drop the tracked coi-serviceworker.js).

   For anyone tracking whether TinyGo makes this release: #522 merged on 2026-07-23 and is inside the range at every cut point under discussion, so it needs no action to be included. #552 is sometimes cited alongside it and is a different change, a bench-suite rework that shipped in v1.12.0 and is contained in v1.12.2.

   The limitation text we would ship with it is wrong. The #504 re-run on 2026-09-06 established that the guide's xxh3 bullet describes a substitution that does not occur: pkg/rt/interop_xxh3.go is !tinygo-gated, so a TinyGo build has no xxh3 namespace and a call fails at the site rather than returning a different digest. The 64-bit int boundary also binds more narrowly than documented, because hash_murmur3.go is untagged and masks to 31 bits; with the seed hash swapped, xsofy under TinyGo wasm produces a world byte-identical to the stock-Go wasm build at a fixed seed.

   PR #799 carries the correction and is open. If the notes end up advertising the TinyGo lane, I would want it in the cut alongside #800, so we are not describing a headline feature incorrectly.

   Two arcs I never gave a section

   Interop, now partly in the cut. #773 merged on 2026-09-06, putting self-contained interop packages and an importable pkg/cli inside the range with no section describing them. #776 (multi-return boxing), #778 (native collection crossing), and #787 (JVM string and keyword surface) are approved and open; #775, #777, and #790 track the rest. Whether this is a headline arc or a footnote now depends only on how long the tag waits.

   Clojure divergences. #761 (reify could not override Object methods) closed on 2026-09-06 when #762 merged, so that one is in the range. Still open: #763 (small maps don't preserve insertion order; PR #764 has changes requested), #742 (nth accepts maps and sets, indexed?/counted? invert for seqs), and #801 (reader metadata yields a with-meta list, map-interior #_ discard fails). #771, which catalogs the known divergences, merged on 2026-09-06 and is in the range.

   Checklist status

   Unchanged since 2026-09-02 apart from the #545 item.

   On the version: v1.13.0 still looks right to me for either candidate cut, since the range already carries the #604 Go API removal alongside features, and the bundle format version bump makes the case for a minor rather than a patch. That holds whichever SHA we freeze at.

   What I am carrying, none of which depends on the timing:

   • #796 and #721 are mine to fix. #796 is a correctness bug in what --entry-frame emits, so I would not want the notes advertising the feature ahead of it. #783 and #782 sit on the same surface.
   • #791 is open on the perf side; I will not write a perf paragraph until it resolves and I have a pinned SHA to measure against.
   • #799 carries the TinyGo guide correction, so it is worth having in before the notes describe that lane.
   • The bundle format entries above want folding into the behavior-change section whenever the body gets rewritten, since that work has already merged.

   Happy to take the body rewrite once the wave settles.

   👀React with 👀1Reacted by mg
6. mparrett commented on Sep 9, 2026

   

   mparrett

   on Sep 9, 2026

   CollaboratorAuthor

   More actions

   Delta, 2026-09-09

   main @ a9c183f9. Nine PRs merged since the note above, and one number in it needs reading differently.

   The items I said I wanted in have landed. #799 (TinyGo guide correction) and #800 (-strip composing with -w) merged on 09-07, #683 on 09-09, and the approved interop set #776 / #778 / #787 on 09-06. That last one settles the question I left open: interop is fully in the range now, not partly, so it wants its own behavior-change section rather than a footnote. The range is 114 squash-merged PRs, up from 105.

   My insertion count was misleading. I quoted 73,075 insertions, and roughly 40% of that was docs/perf/baseline.json, which #684 then reseeded from a window (30,726 lines down to 8,699). Excluding that file the range is +48,412 / -8,647 across 406 files, and that is the figure worth using for release-notes sizing. The clearest sign the raw number is not measuring code: it fell from 73,075 to 55,511 over three days in which nine PRs merged and the code grew.

   One behavior change to add to the list. #764 merged on 09-07 and closed #763: small maps now preserve insertion order, matching Clojure array-map semantics. #826 is open to narrow the divergence-catalog entry to construction. That is consumer-visible and belongs with #758 / #760 / #756 in the notes.

   Unchanged: #796, #791, #783, #782, and #721 are all still open, so the entry-frame set and the perf question stand where they did. Still no preference from me on timing.
7. mparrett commented on Sep 19, 2026

   

   mparrett

   on Sep 19, 2026

   CollaboratorAuthor

   More actions

   Body rewritten, 2026-09-18

   main @ 7148f4ad. I said on 09-06 that I would take the body rewrite once the merge wave settled. It has: thirteen PRs are open with three approvals, against twenty open and four approved in early September. So the arcs can be written from the whole range instead of layered as another delta on a stale snapshot.

   The body is now a consolidated statement of capabilities, bugfixes, compatibility, performance and open items, rather than the 2026-08-11 text plus three correction comments. Those comments stay where they are; nothing in them is lost, and the corrections they carry are folded in. Where a comment disagrees with the body, the body is current.

   Fifty PRs merged since the 09-09 delta, taking the range to 164 squash-merged PRs plus three direct commits across 167 commits, +64,157 / -9,398 across 540 files excluding docs/perf/baseline.json.

   What those fifty add that the earlier notes had no section for:

   • An HTTP and networking arc. http/start, http/stop and http/wait (http: stoppable server, empty :headers fix; core: & rest binds nil #898), streaming response bodies (feat(http): stream channel and lazy-seq http/serve response bodies (#831) #851), three timeout scopes (feat(http): connect, request and stream_read timeouts for http/request (#844) #856), scope cancellation (fix(http): http clients inherit the caller's scope cancellation (#816) #848), and TCP listeners through net/listen, net/accept and net/local-address (feat(rt): net/listen, net/accept and net/local-address — owned TCP listeners #896).
   • A Clojure-surface arc. Custom data readers (feat(reader): support custom data readers #770), a live *unchecked-math* (feat(compiler): honor *unchecked-math* for + - * inc dec #839), the Math/* statics (feat(rt): add missing JVM numeric statics (Math/*, Long/bitCount, ...) #840), canonical (. obj member ...) forms (feat(compiler): accept Clojure's canonical (. obj member ...) forms #841), clojure.edn data semantics (fix(edn): clojure.edn reads data with Clojure data semantics (#801, #823) #854), and Pattern.split semantics for clojure.string/split (fix(string): clojure.string/split follows Java Pattern.split semantics #843).
   • 64-bit Int on every host (fix(vm): make Int 64 bits wide on every host #859), with xxh3 reaching TinyGo wasm (fix(rt): register the xxh3 namespace on TinyGo wasm builds #860) and the lowerer typing :int as int64 (fix(ir): lower a let-go int to Go int64 on every host #862). This is what makes a program portable across the two widths; the residual boundaries are Wide Int values narrow at collection, math, and interop boundaries on 32-bit hosts #867.
   • lg.compiler embedded as a core namespace (feat(aot): embed the AOT compile driver as the lg.compiler namespace #735), so a shipped lg drives AOT lowering from any directory, plus unchecked-* as first-class IR and VM ops (feat(ir,vm): unchecked-add/subtract/multiply as first-class ops #811).
   • glplat merged behind build tags (glplat: Ebitengine backend behind -tags glplat_ebiten (stacked on #392) #744), with untagged builds gated against it (build: gate untagged builds against the optional heavy subsystems #789).

   Three items for the cut specifically:

   1. Two independent bundle-compatibility breaks, not previously written down together. feat(ir,vm): unchecked-add/subtract/multiply as first-class ops #811 changes the opcode-set signature, so an older lg's .lgb is rejected; feat(bytecode): opt-in DEFLATE compression for .lgb bundles #501's compressed bundles are format version 3, which an older decoder rejects as unsupported. Both are the intended design, and both read as corruption if the notes do not say so.
   2. [x & more] now binds nil, not () (http: stoppable server, empty :headers fix; core: & rest binds nil #898). The old behavior hung integrant's ig/init on any three-component chain, so this is a fix, and it is also a visible semantic change that belongs in the notes with feat(vm): insertion-ordered small maps matching Clojure array-map semantics #764, feat(vm): symbols are invokable for collection lookup, like keywords #758, fix(vm): multi-arity dispatch routes exactly-min-arg calls to the variadic variant #760 and fix(compiler): var special form resolves private vars across namespaces #756.
   3. The perf section carries no numbers, deliberately. bench-ratchet: IRCompile and InitFromLGB regressed since the #780 baseline (#727 allocs, #723/#730 timing) #791 is open against the build: bench-ratchet infrastructure — deterministic rebaseline, Go 1.26.5 baseline, pre-push gate #780 baseline, and bench-ratchet: deterministic bars come from profiles captured a month apart #883, bench-ratchet: gate timings on an interval from the stored samples, not a point against a 5% budget #899, pkg/vm: BenchmarkConsCreation measures an empty loop #884 and pkg/vm benchmarks name the same size two ways: /1000 and /01000 #885 are four further gate defects filed since. The audit should be one pass against a pinned SHA, not a table anchored to main.

   The only item I would call a blocker is #796: entry-frame binaries run -main twice, and the notes plan to advertise --entry-frame. Either that lands or the flag stays out of the advertised surface. #783, #782, #721 and #866 sit on the same feature.

   Still no preference from me on timing.
8. mparrett commented on Sep 19, 2026

   

   mparrett

   on Sep 19, 2026

   CollaboratorAuthor

   More actions

   Delta, 2026-09-19

   main @ 827ef18b. Five PRs merged since yesterday's rewrite — #894, #795, #863, #871, #900 — taking the range to 169 squash-merged PRs plus three direct commits across 172 commits. I have refreshed the body; this comment covers what changed in it and three items that need someone else's hand.

   #863 replaced #852's mechanism, so the body's compatibility note was describing code that no longer ships. #852 rebuilt a bound-fn* call's context from the captured snapshot alone, which silently reverted any dynamic var the caller had bound but that was not captured at creation. Since *file* is bound while loading a file, the snapshot is essentially never empty, so that was the path taken nearly always. #863 overlays the captured vars onto a private child of the caller's context instead, matching Clojure's push-thread-bindings, and keeps the property #852 fixed.

   That resolves #864, which is still open because #863 never referenced it. I verified rather than assumed, with a failing control:

   (def ^:dynamic *x* :root-x)
   (def ^:dynamic *y* :root-y)
   (let [f   (binding [*x* :captured]     (bound-fn [] [*x* *y*]))
         got (binding [*y* :caller-bound] (f))]
     got)

   9d32c1c4 (post-#852, pre-#863) returns [:captured :root-y]. 827ef18b returns [:captured :caller-bound]. My first attempt at this repro created the bound fn with an empty capture snapshot and passed on both builds, which is the reminder that a green check against one tree proves nothing without the control. #864 looks closeable.

   #900 resolves #883, also without referencing it. #900 takes the deterministic bar per benchmark from the entry with the newest provenance instead of the global minimum across profiles captured at different commits, stamps provenance in seed-baseline, and names the source commit in the regression line. That is #883's own options 1 and 3, and #900's write-up reproduces #883's exact evidence: the 8370C August row holding 7,041 allocs/op for InitFromLGB against 13,995–26,428 in every newer row, gating current main at +69%. #883 refs nothing and stays open; it should be closed. With #795's bench-baton lease alongside it, gate repair is now ahead of the defect list for the first time in this cycle.

   #863 merged ahead of its own spec. Its commit subject cites docs/specs/clojure-test-conformance.md, and that file is not on main — it is #793, still open with changes requested. docs/specs/ currently holds only executable-evidence.md. Worth landing #793 before the tag, since the squash subject is permanent.

   Also folded into the body: clojure.test is now a real port (report, assert, fixture, runner) rather than the handful of additions the Clojure-surface paragraph listed; #894 published the binary-size and boot-time audit behind #652, which is the method and tooling for the checklist's audit item but not the audit of this cut, since it compares v1.7.4 against the historical ed4ecc2; and #906 takes the pkg/rt leaf conversions from #867, leaving the pkg/vm half until the ToInt question is settled.

   Unchanged: #902 has not been reviewed, so #796 is still the one blocker, and nothing has frozen the scope — the range grew by five PRs in a day.
9. mparrett commented on Sep 19, 2026

   

   mparrett

   on Sep 19, 2026

   CollaboratorAuthor

   More actions

   Release notes, Option A — the sober version

   Two drafts for the v1.13.0 notes, so nobody starts from a blank page. This is the straight one, for the GitHub release body. Option B in the next comment is the same release written as a Slack post in @nooga's voice.

   Both are starting points to cut and rewrite, not proposals. Measured at main @ 827ef18b, and written assuming #901 and #902 land.

   ---

   v1.13.0 is the largest release let-go has cut: 169 merged pull requests in the 61 days since v1.12.2, against 126 for the next biggest range in the project's history. Two changes in it move what you can build, and a third moves where you can run it.

   A let-go program now computes the same values on every host. Int was type Int int, so its width followed the platform. On linux/386, linux/arm and TinyGo wasm, (bit-shift-left 1 40) returned 0, (* 1e9 1e9) trapped, and a uint64 crossing an interop boundary was truncated. It is 64 bits wide everywhere now, the Go lowerer types :int to match, and xxh3 reaches TinyGo wasm for the first time. A 64-bit hash chain produces byte-identical results on a small-footprint wasm build and on amd64, which is the property that makes those targets usable for anything that has to agree with itself.

   The HTTP namespace became something you can run a service on. http/serve was ListenAndServe on the calling goroutine: no handle, no stop, a bind error the only way back. It is now http/start, http/stop and http/wait, with channel and lazy-seq response bodies flushed element by element, three named timeout scopes on http/request, and clients that inherit the caller's structured-concurrency scope so cancelling a scope aborts the requests under it. net/listen, net/accept and net/local-address give TCP the owned-listener surface unix already had. The forcing function was real code: an integrant and SQLite web app built under lgx, and the nREPL hub that had been carrying a downstream patch until now.

   Real Clojure libraries drove the compatibility work, and they name themselves in the diff. weavejester/dependency exposed [x & more] binding an empty seq instead of nil, which hung integrant's ig/init on any three-component chain. clojure.test.check wanted *unchecked-math* to stop being inert and the Math/* statics to exist. clojure.data.priority-map wanted Clojure's canonical (. obj member ...) forms to compile. weavejester/ragtime wanted Thread/currentThread, which now resolves to the calling scope, so the cooperative-cancellation idiom compiles and .isInterrupted means scope cancellation. Around them: small maps preserve insertion order, reify overrides Object methods, symbols are invokable for lookup, clojure.edn reads data rather than code forms, and clojure.string/split follows Pattern.split. The divergence catalog shrank by most of its consumer-visible entries. clojure.test was ported outright — a real report, assert, fixture and runner contract rather than the partial namespace that preceded it.

   AOT stopped needing a source checkout. The lowering driver is now the embedded lg.compiler namespace, so a shipped lg drives lowering from any directory, and lg-compile --entry-frame emits a native-entry Go frame from main or -main. unchecked-add, unchecked-subtract and unchecked-multiply became first-class IR and VM ops, so a hash or PRNG inner loop keeps its typed native path instead of boxing at every step. On footprint: opt-in DEFLATE bundles, lg -strip splitting debug info into a digest-bound companion rather than discarding it, lg_no_http to drop net/http from the runtime, and the embedded core bundle down from 311,302 to 273,748 bytes.

   What it costs. Two independent bundle-compatibility breaks — the opcode set changed, and compressed bundles are format version 3 — so bundles do not cross this boundary in either direction. Both rejections are by design, and both read as corruption if the notes do not say so. The toolchain floor is Go 1.26.5. Two Go API moves: 34 per-file core-embed symbols are gone, and the CLI moved out of the root package into an importable pkg/cli. Scope cancellation now propagates where it previously did not, a correction that will stop code relying on requests outliving their scope. On performance the direction is good — reduce measures below the v1.12.2 baseline rather than merely recovered toward it — but the measurement apparatus is under active repair, so the release should carry one audit against a frozen SHA rather than numbers anchored to a moving tip.

   ---

   One caveat for whoever finalizes this

   Written assuming #901 and #902 land. If either slips, pull its line: #901 is Thread/currentThread and the %n fix, #902 is what makes --entry-frame safe to advertise.

   glplat is experimental with an unstable contract, per @nooga's call recorded in docs/design/glplat-backend-contract.md §5.1. The package is marked provisional: the registry stays under internal/, and the planned rt.Display follow-up may rename natives and change Init's signature. The namespace is present in a default build with every native returning no backend registered, and no graphics dependency linked. Name it experimental or leave it out of the summary; do not describe it as a stable surface.

   ❤️React with ❤️2Reacted by dharmatech and mg
10. mparrett commented on Sep 19, 2026

    

    mparrett

    on Sep 19, 2026

    CollaboratorAuthor

    More actions

    Release notes, Option B — the announcement, in @nooga's voice

    The same release shaped as a Slack post, matching the register of your v1.12.1 announcement: the #let-go opener, the bullet blocks, the ecosystem roundup, brew install, the star ask.

    Written in your first person because that is the form it ships in — a draft for you to take, cut, or rewrite, not something anyone else should post as-is. The notes at the bottom cover what I left out and why.

    ---

    let-go v1.13.0 is out 🎉

    #let-go is a Clojure implemented in Go, a bytecode compiler and VM, with an AOT compiler written in let-go itself. One ~14MB binary.

    This is the biggest release we've ever cut: 169 merged PRs in the two months since v1.12.2, more than any release in the project's history — the previous record was 126.

    The same program now computes the same numbers on every host.

    • Int used to be Go's int, so its width followed the platform. On 32-bit targets — linux/386, linux/arm, TinyGo wasm — (bit-shift-left 1 40) quietly returned 0 and (* 1e9 1e9) trapped. • It's 64 bits everywhere now. The AOT lowerer types to match, and xxh3 works on TinyGo wasm for the first time. • xsofy now generates a byte-identical world from the same seed on a TinyGo wasm build and on native — @matt Parrett chased that one down through its determinism layer. That's the property that makes the small targets worth shipping to.

    You can write a web service in let-go now.

    • http/serve used to be ListenAndServe on your goroutine: no handle, no stop. Now there's http/start, http/stop and http/wait. • Response bodies can be a channel or a lazy seq, flushed element by element — SSE and long-polling work. • Three named timeout scopes on http/request: connect, request, stream_read. • Clients inherit the calling scope, so closing a scope cancels the requests under it. • net/listen, net/accept and net/local-address — TCP finally has the listener side, from @norman Nunley's patch that had been living downstream.

    More real Clojure libraries run, unmodified.

    • integrant works. [x & more] was binding an empty seq instead of nil, which hung weavejester/dependency's reachability loop, so ig/init deadlocked on any three-component chain. That one took a while to find. 😅 • clojure.data.priority-map compiles — we accept Clojure's canonical (. obj member ...) forms now, not just the .member sugar. • clojure.test.check groundwork: *unchecked-math* is live instead of inert, and the Math/* statics exist, each pinned to JVM semantics rather than Go's. (Java rounds -2.5 to -2. Go doesn't. Now we do.) • weavejester/ragtime loads — Thread/currentThread resolves to the calling scope, so (.isInterrupted (Thread/currentThread)) compiles and actually means "was my scope cancelled". format honors %n too. • clojure.test is a real port now — report, assert, fixture and runner — not the handful of functions we had been getting by with. • Divergences closed: small maps keep insertion order, reify overrides toString/hashCode/equals, symbols are invokable for lookup, clojure.edn reads data rather than code forms, clojure.string/split follows Pattern.split, and custom *data-readers* work.

    The AOT compiler has a front door.

    • Driving lowering used to need a let-go checkout on disk. The driver is now the embedded lg.compiler namespace, so a shipped lg compiles from any directory. • unchecked-add/subtract/multiply became first-class IR and VM ops, so a hash or PRNG inner loop keeps its typed native path instead of boxing at every step. • lg-compile --entry-frame emits a real native entry point, so an AOT binary starts in Go rather than replaying a main chunk.

    Smaller, in several directions.

    • Opt-in DEFLATE for .lgb bundles. • lg -strip splits debug info into a digest-bound companion file instead of throwing it away, and it composes with -w now. • lg_no_http drops net/http from the runtime entirely. • The embedded core bundle went from 311,302 to 273,748 bytes.

    Graphics, experimentally. glplat landed: a platform layer with two backends, GLFW/OpenGL and Ebitengine — the second reaching the browser and Metal where the first is cgo-native-only. The contract is explicitly unstable and the namespace is experimental, so please play with it but don't build on it yet.

    Two things to know before you upgrade:

    • .lgb bundles don't cross this release in either direction. The opcode set grew and compressed bundles moved to format version 3, so rebuild your bundles. Both rejections are deliberate — if you see one, it isn't corruption. • The toolchain floor is Go 1.26.5, 34 unreferenced core-embed symbols are gone, and the CLI moved out of the root package into an importable pkg/cli — so you can now build your own lg with the full feature set around your own interop packages.

    The ecosystem keeps growing.

    • @Andrey Bogoyavlenskiy's lgx is the project and package manager, and a real integrant + SQLite web app built on it is what drove most of the HTTP work above. • @norman Nunley's strange-lettractor runs an nREPL hub and OAuth loopback on let-go. • xsofy runs in the browser, in a terminal, on Plan 9, and now under TinyGo. • legmacs got a better deal than I expected out of this one: eval now runs in the caller's execution context, so its built-in REPL can capture what your form prints, not just what it returns. • @jarkko Saltiola built a GTK time tracker that embeds let-go and hot-reloads its UI over nREPL — which is why api.LetGo exposes its compiler context now.

    Welcome to our two first-time contributors, @jarkko Saltiola and @dharmatech 👋 — dharmatech fixed Plan 9 key reads hanging on terminal resize, which is exactly the kind of bug that only gets found by someone actually running it there.

    Give it a try:

    brew install nooga/tap/let-go

    Release notes: https://github.com/nooga/let-go/releases/tag/v1.13.0 Repo: https://github.com/nooga/let-go (please ⭐ it, we live for ⭐s)

    Thank you to everyone who contributed, filed, reviewed and kicked the tires 🙏

    If you maintain a Clojure library: we'd still love to hear what breaks. Most of the compat work above started as someone trying to load something and telling us what happened.

    ---

    Notes for whoever finalizes this

    • Written assuming host: Thread/currentThread over scopes; core: format honors %n #901 and fix(aot): stop entry-frame binaries from running -main twice #902 land. If either slips, pull its bullet: host: Thread/currentThread over scopes; core: format honors %n #901 is the ragtime line, fix(aot): stop entry-frame binaries from running -main twice #902 is the --entry-frame line in the AOT block. fix(aot): stop entry-frame binaries from running -main twice #902 is what makes that flag safe to mention — without it, entry-frame binaries run -main twice.
    • No perf headline. v1.12.1's announcement led with fib 2.42s → 0.11s, and there's no equivalent honest number here: bench-ratchet: IRCompile and InitFromLGB regressed since the #780 baseline (#727 allocs, #723/#730 timing) #791 is open against the build: bench-ratchet infrastructure — deterministic rebaseline, Go 1.26.5 baseline, pre-push gate #780 baseline, with four more gate defects behind it. The one defensible figure is reduce at 50.5 ms against 58.7 ms on v1.12.2 — below the old baseline, not merely recovered. Worth a line if the audit confirms it at the frozen SHA.
    • The library list is shorter than v1.12.1's on purpose. That announcement could name HoneySQL, medley, hiccup, malli, integrant, aero and tools.cli as running. I only re-verified the three this release specifically fixed, so the rest should either be re-checked or carried forward with a "still runs" framing rather than re-listed as news.
    • The contributor stats behind the credits: mparrett 91 PRs, nnunley 59, abogoyavlensky 18, jasalt 3, nooga 3, dharmatech 2.

    What's in this release for legmacs

    Checked against the working tree at 2b25d9b. Three findings, one of which is a trap rather than a win.

    legmacs is pinned behind all of it. .let-go-version is c879db76 (2026-07-01), which predates v1.12.2 by nearly three weeks and sits 294 commits behind main. None of what follows arrives until that pin moves, and the file's own comment says to bump it "when you want newer let-go fixes".

    #853 is the real win, and it lands on the feature legmacs is built around. legmacs/modes/letgo.lg is an in-editor REPL: eval-last-sexp and eval-print-last-sexp call (eval form) at line 229, and the namespace header reasons explicitly about what context eval runs in. Before #853, eval ran its frame against the root execution context, so dynamic bindings around it did not apply. After it, they do — which means wrapping that call in with-out-str captures what the user's form prints. Right now eval-last-sexp can only show a return value; (println …) output goes nowhere the editor can see. That is a small patch on their side and a visible feature for anyone using the REPL mode.

    #735 could take the source checkout out of their build. build.sh requires a let-go source tree (LETGO_DIR, defaulting to ../let-go), builds lg from it, then runs "$lg_bin" scripts/lg-compile. The pin comment gives the reason: the AOT tooling "lives on main". After #735 the driver is the embedded lg.compiler namespace, so a released lg drives lowering from any directory, and scripts/lg-compile survives as a shim with its path, argv and exit codes unchanged. Their existing invocation keeps working; the checkout requirement is what goes away.

    #843's split-lines is not a drop-in for theirs — don't let anyone swap it. legmacs/buffer.lg:10 hand-rolls split-lines with a documented editor contract: one trailing newline does not produce a phantom empty last line, but any other newline does. let-go's new clojure.string/split-lines is (split s #"\r?\n") at limit 0, which per Pattern.split drops all trailing empty fields. The two agree on "" and on a single trailing newline, which is exactly what test/buffer_test.lg pins, so a swap would pass their tests and silently eat the blank lines at the end of a file. The \r\n handling is worth borrowing; the function is not.

    ❤️React with ❤️2Reacted by dharmatech and mg
11. nooga commented on Sep 20, 2026

    

    nooga

    on Sep 20, 2026

    Owner

    More actions

    Checked off three checklist items and a status update, following on from the 2026-09-19 refresh.

    Seven more PRs merged since 827ef18

    All reviewed, tested and landed 2026-09-19/20:

    • host: Thread/currentThread over scopes; core: format honors %n #901 — Thread/currentThread resolves to the calling scope; format honors %n.
    • fix(aot): stop entry-frame binaries from running -main twice #902 — fixes entry-frame binaries run -main twice: main-chunk replay, then the frame #796 properly: lg -c now omits only the selected top-level entry call from the frame-specific bundle, instead of gating *compiling-aot* across the whole replay (which was dropping required-namespace runtime init — the semantics bug @nooga flagged in review). Verified with the exact two-namespace repro from that review thread.
    • fix(rt): keep 64-bit integers across the rt boundary on 32-bit hosts #906 — 64-bit ints survive the pkg/rt boundary on 32-bit hosts (the leaf-conversion half of Wide Int values narrow at collection, math, and interop boundaries on 32-bit hosts #867; pkg/vm's ToInt half is still deferred).
    • ci: move the TinyGo gate to 0.42.0 #909 — TinyGo gate moved to 0.42.0.
    • build: bump Go to 1.27 and pin golangci-lint in go.mod #910 — Go bumped to 1.27, golangci-lint pinned alongside it in go.mod.
    • fix(test): expand one-arity is directly to try-expr #915 — one-arity is no longer recurses through its own unqualified name; fixes 47/242 jank-suite files that were silently failing to compile since feat(test): port clojure.test — real report/assert/fixture/runner contract #863 (compile failures were t.Skip'd, so CI stayed green through the loss).
    • fix(genmanifest): pin the manifest toolchain from go.mod, not GOTOOLCHAIN=local #916 — the manifest generator now pins GOTOOLCHAIN from go.mod's toolchain line instead of GOTOOLCHAIN=local, so a contributor on an older Go can still make generate (this is what made build: bump Go to 1.27 and pin golangci-lint in go.mod #910 safe to use locally — hit this firsthand while sanity-checking the cut).

    Current main: 369e2a6900a4.

    Checklist

    • entry-frame binaries run -main twice: main-chunk replay, then the frame #796 — closed, fixed by fix(aot): stop entry-frame binaries from running -main twice #902 (above). Checked.
    • bound-fn*: uncaptured dynamic vars revert to root inside the call (#852 replaces the binding stack instead of pushing onto it) #864, bench-ratchet: deterministic bars come from profiles captured a month apart #883 — already closed (2026-09-19, by nnunley and mparrett respectively). No action needed.
    • Release hook — verified 2026-09-20 against current main: make check-generated, make check-generated-manifest, make build, make lint, go test -short ./... all pass clean. Checked.
    • Upgrade notes — drafted (not published) to RELEASE_NOTES_v1.13.0.md, organized by theme rather than a flat PR list, cross-referenced against this issue's own checklist. @nooga is reviewing/editing before anything is posted. Checked as "drafted"; still needs the actual publish step.

    Still open, unchanged

    • Three gaps that let the perf gate report green through a regression #705, bench-ratchet: gate timings on an interval from the stored samples, not a point against a 5% budget #899 — bench-ratchet gate defects, undecided.
    • AOT lower-vm: embedded VM-fallback bundle rejected by its own runtime — unsupported capability mask 0x1 (supported: 0x0) #607, strict-mode AOT: 6/16 gold-aot fixtures only run via silent fallback — binding conveyance, @f deref rot, closure capture untracked #678 — AOT matrix gaps (lower-vm capability rejection, strict-mode fallback), undecided.
    • docs(specs): specify clojure.test conformance, clojure.test.tap, and stack-trace primitives #793 — clojure.test-conformance.md spec still doesn't exist on main; feat(test): port clojure.test — real report/assert/fixture/runner contract #863's commit message still points at it.
    • Size/perf audit vs v1.12.2 at a pinned SHA — not run. docs(perf): publish the binary-size and boot-time audit behind #652 #894 is the method, not this range's audit.
    • Scope freeze — not formally declared. Given the remaining items above are all pre-existing/deferred work rather than something this range's PRs touch, 369e2a6900a4 looks like a reasonable freeze candidate, but that call is @nooga's.

    New during this pass

    • TestJankSuiteDirectABIGeneratedGo is red on main since #863, and no CI lane runs it #923 — TestJankSuiteDirectABIGeneratedGo red on main since feat(test): port clojure.test — real report/assert/fixture/runner contract #863 (bisected), separate from what fix(test): expand one-arity is directly to try-expr #915 fixed. No CI lane currently runs it. Filed by mparrett; open question for @nnunley on whether the new AST is correct-by-design or a real regression. Not release-blocking (skipped under -short), but worth a decision before or shortly after the tag.
12. mparrett commented on Sep 20, 2026

    

    mparrett

    on Sep 20, 2026

    CollaboratorAuthor

    More actions

    Correction for the two release-note drafts above: the Go floor moved

    Heads-up before anything gets published — both Option A and Option B say Go 1.26.5, and that is now wrong. #910 bumped the module to Go 1.27 with toolchain go1.27.1 and pinned golangci-lint alongside it, several hours after I posted those drafts.

    The line to change, in both:

    The toolchain floor is Go 1.26.5

    The toolchain floor is Go 1.27

    Worth pairing it with #916 in the upgrade notes, since the two travel together: the manifest generator now takes GOTOOLCHAIN from go.mod's toolchain line instead of GOTOOLCHAIN=local, so a contributor on an older local Go can still run make generate. Without that, the bump would have been a harder break for anyone not already on 1.27.

    The toolchain floor is the most likely line in the whole set to be acted on by a downstream consumer, so it seemed worth flagging on its own rather than folding into the next refresh.

    The issue body is already updated: Go 1.27, and the range restated at 369e2a69 — 176 squash-merged PRs plus three direct commits across 179 commits, +81,395 / -9,772 over 659 files excluding docs/perf/baseline.json. The counts in the two drafts (169 PRs, 61 days) are one day stale for the same reason; they read fine as "the two months since v1.12.2" if you would rather not chase the number.
13. mparrett commented on Sep 27, 2026

    

    mparrett

    on Sep 27, 2026

    CollaboratorAuthor

    More actions

    v1.13.0 shipped

    v1.13.0 was tagged on 2026-09-20 at 369e2a69, the freeze candidate the body names, so the scope described above is the scope that shipped. The release notes carry the upgrade notes, the bundle-compatibility breaks, the Go 1.27 floor, and glplat marked experimental.

    Checklist outcome

    Item	Outcome
    Freeze at a named SHA	369e2a69, the tagged commit
    #545	Closed 2026-09-02
    Size and perf audit vs v1.12.2	Not run. The notes say so and ship no headline number. #894 is the method, not this range's audit
    #705, #883, #899	#883 closed by #900. #705 and #899 shipped undecided and are still open
    AOT matrix	--entry-frame advertised after #902. #607 and #678 are still open, as are #783, #782, #721 and #866 on the entry-frame surface
    #796	Fixed by #902
    Release hook	Green on 369e2a69 (@nooga, 2026-09-20)
    Upgrade notes	Published in the release body
    Release and attribution	Published, with contributor credits
    Tag	v1.13.0

    Two items the body wanted settled before the tag are still open:

    • docs(specs): specify clojure.test conformance, clojure.test.tap, and stack-trace primitives #793. docs/specs/clojure-test-conformance.md is still absent from main, so feat(test): port clojure.test — real report/assert/fixture/runner contract #863's squash subject cites a file that does not exist. docs(specs): specify clojure.test conformance, clojure.test.tap, and stack-trace primitives #793 is in changes-requested.
    • TestJankSuiteDirectABIGeneratedGo is red on main since #863, and no CI lane runs it #923. TestJankSuiteDirectABIGeneratedGo is still red, and still run by no workflow.

    Found after the tag

    Against the shipped release:

    • (run-tests) with no arguments silently runs zero tests after #863 #929. A bare (run-tests) now runs only the current namespace, which is Clojure's behavior and the point of the feat(test): port clojure.test — real report/assert/fixture/runner contract #863 port, but an aggregate runner that requires its test namespaces and calls (run-tests) now reports zero tests and exits 0. The release notes present the port as a feature and do not list this under Breaking. A one-line addition there, pointing at (run-all-tests #"myproj\..*"), would cover it.
    • lg -c/-b: lazily-loaded core namespaces bundle with every var nil #954. A program that requires a lazily-loaded core namespace (zip, data, check) runs from source and fails from its .lgb with TypeError: nil is not a function. It predates this release: the same repro fails on both the v1.12.2 and v1.13.0 binaries. fix(rt): load embedded core namespaces a compiled program requires #966 fixes it and is approved.
    • One file cannot serve two let-go versions: no read-time capability discriminator #938. A file cannot serve two let-go versions, because an unknown symbol is a compile error even in an untaken branch. Surfaced by lgx test: run on let-go's clojure.test port (#863) while keeping lg 1.12 working abogoyavlensky/lgx#52 against the clojure.test port.
    • AOT wasm (LG_WASM_BUILD_TAGS=gogen_ir) cannot build from a normally-built lg #940. LG_WASM_BUILD_TAGS=gogen_ir lg -w cannot build from a normally-built lg.
    • glplat: open review findings from #392, now in main via #744 #931. The review findings from glplat: minimal GL platform layer (experiment) — GLFW/OpenGL backend, font primitives, screenshot #392, which closed on 2026-09-20, are now tracked against the glplat code that shipped via glplat: Ebitengine backend behind -tags glplat_ebiten (stacked on #392) #744.

    The let-go-wiki refresh for this release is done (nnunley/let-go-wiki#28).

    Carried forward

    For whoever opens the next tracking issue: the size and perf audit at a pinned SHA, #705 and #899, the AOT matrix (#607, #678, and the four entry-frame issues above), #793, #923, #867's pkg/vm half, and #929's release-notes line.

    Closing, since everything left is tracked in its own issue.
14. mparrett

    closed this as completedon Sep 27, 2026