Skip to content

LSP on lspkit 0.10.1; drop unused dependencies - #417

Merged
MelbourneDeveloper merged 83 commits into
mainfrom
chore/lspkit-0.10
Nov 20, 2026
Merged

MelbourneDeveloper merged 83 commits into
mainfrom
chore/lspkit-0.10

Conversation

@MelbourneDeveloper

Copy link
Copy Markdown
Collaborator

TL;DR

Moves the language server to lspkit 0.10.1 and drops dependencies nothing uses.

Details

  • lspkit 0.6.0 → 0.10.1: lspkit, lspkit-server and lsp-profiler in src/toolchain/Cargo.toml, plus @nimblesite/lsp-profiler-ui ^0.10.0, which is its newest release.
    • The releases in between fix leaks and lifecycles: an abandoned cooperative attach now stops the in-process sampler, forwarder tasks are owned and pinned, Python memory tracking is released on stop, and decoration types are disposed.
    • 0.10.1 also fixes the frame drain (lspkit Spec Cleanup #55) and logs dropped sidecar notifications, which adds a tracing edge.
    • Osprey uses no API that changed. lsp-profiler still pins sysinfo 0.39.3, which matches the workspace.
  • Unused dependencies removed:
    • thiserror from lsp-profiler-osprey.
    • ts-node and its .vscodeignore line from the VS Code extension.
    • @11ty/eleventy-plugin-syntaxhighlight and markdown-it-anchor from the website.
    • @osprey/graphics and @osprey/json from signal_flow and vscode, and @osprey/json from uitest. These packages never import them; @osprey/native brings them in.
  • Kept: the website's markdown-it devDependency. tests/pages.spec.js calls require("markdown-it"), which fails with MODULE_NOT_FOUND once the package is gone from the top level.
  • docs/plans/lspkit-leak-upgrade.md now records the pinned version.

How Do The Automated Tests Prove It Works?

  • The LSP suites run on lspkit 0.10.1: cargo test -p osprey-lsp and the live_profiler_lsp end-to-end test, which drives lspkit's profiler service and covers $/cancelRequest. The VS Code extension's LSP suite runs as well.
  • The website test long authored guides retain a useful image or inline video per 500 prose words passes with markdown-it restored. Without it, it fails with MODULE_NOT_FOUND.
  • The package matrix test the_package_matrix_runs_all_62_suite_backend_pairs_once builds signal_flow, uitest and vscode without the removed manifest entries.
  • Local make ci on macOS: lint, build and every test pass. The three Rust coverage gates that fail only on macOS (osprey-cli, osprey-lsp, osprey-types) are the same ones that failed locally for fix(lsp): re-check only affected projects (#384); lspkit 0.6.0 #413, whose Linux CI passed them.

🤖 Generated with Claude Code

- lsp-profiler-osprey: thiserror (no use).
- website: markdown-it, markdown-it-anchor, @11ty/eleventy-plugin-syntaxhighlight
  (only eleventy-plugin-techdoc uses them, through its own dependencies).
- vscode-extension: ts-node (no use) and its .vscodeignore line.
- signal_flow, vscode: @osprey/graphics and @osprey/json; uitest: @osprey/json.
  They come in transitively through @osprey/native and the packages never name them.
test-integration ran the bank, profiler, agent, Playwright, docs and web
compiler suites plus the Android emulator under one 18-minute limit; green
runs took 14-18 minutes and PR 417 timed out twice. test-android now runs
the emulator with the same limit. Both jobs pull the LFS phone golden:
the emulator compares against it and test-mobile-tools.py decodes it.
Ubuntu 26 runners no longer ship Playwright's browser libraries, and the
Azure mirror served them slowly enough to time the Website E2E job out at
8 minutes. Every apt install in the job now keeps its packages in a cache
that is saved right after the installs, and WebKit installs alongside
Chromium so tests/install-webkit.cjs finds it present.
The image's newer CsWinRT analyzer rejects collection expressions whose
target is a read-only interface (CsWinRT1032: no concrete type for
trimming and AOT), and warnings are errors, so the WinUI host stopped
building with no source change. The 15 flagged sites now build an
explicit array.
The apt cache could never fill: on a cold cache the serial downloads
from the runner's Azure mirror outlasted the 8-minute job, so the save
step never ran. scripts/ci/apt-prefetch.sh asks apt for the archive
URIs and fetches them 16 at a time with curl --parallel; apt then
installs with nothing left to download, checking every archive's hash
as usual. In an Ubuntu 26.04 container it fetched 174 archives (115 MB)
in 43 s and apt reported 'Need to get 0 B'.
…g it

The release gate guard rejected '|| true' after the dependency dry run.
The dry run exits 1 when it lists missing packages; the step now accepts
that one answer and fails on any other.
MelbourneDeveloper and others added 28 commits November 20, 2026 11:42
…t E end to end

Error is only the Result constructor. Fallible built-ins return
Result<T, string>. Error { message } builds and binds the payload at
the Result's own E on GC, ARC and wasm32. expect, check and expectError
are generic over their operand types. Implements [ERR-RESULT] [ERR-PAYLOAD].
…ent; assertions compare one type

Corpus and rejection programs cover typed error payloads (record, union,
int, bool, float, nested Result) through match, ?:, toString and closures.
Implements [TESTING-BUILTIN-EXPECT-ERROR] [TESTING-EQUALITY] [ERR-PAYLOAD].
… [BUILTIN-LIST-GET]

listGet and xs[i] return Result<T, IndexOutOfBounds>; the record is a prelude
unit attached beside every list read, built by the backend at the bounds check
on GC, ARC and wasm32. An index on a not-yet-typed target is a deferred
relation resolved to its container's result. Programs and packages that
relied on the old string message now state their own.
CI run 44498694913 captured these after 2440c10 routed MIDI notes into
oscillator inputs; the old goldens predate the MIDI port and caught the
inspector mid-animation.
…BUILTIN-PRINT-FORM]

A record prints as { code: 1, why: "empty" }, a union by its variant, a list
by its elements, at any depth and identically on GC, ARC and wasm32. A string
inside one is quoted. Functions, maps and runtime handles stay unprintable, and
an opaque alias stays sealed outside its module [MODULES-OPAQUE-TYPES].
…lasses [TYPE-CONSTRAINTS]

Planned, not implemented. The plan ranks it and nested constructor patterns
first, above every other open item.
…iveness [PATTERN-NESTED]

Planned, not implemented; the plan lists the steps.
…and the any-box null guard

Brings deslop back under its 4.00% cap (3.995%). Every folded test keeps its
checks and matches its golden under GC and ARC; the nested Range trace stays
nested.
…arm binds it [BUILTIN-LIST-GET]

3bed6a8 built the record on every failed lookup in a fused listGet match,
so an Error arm that ignored it still allocated per miss. The record is now
built only for an arm that binds the payload; the lowering test reads its
fields and pins zero allocation for unbound and wildcard arms.
…[BUILTIN-LIST-GET]

ML twins spell an ignored failure `Error _message`, which binds a name the
body never reads; it now lowers exactly like `Error`, so ML and Default twins
emit identical IR again.
…IN-PRINT-FORM]

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…iler [BUILTIN-PRINT-FORM]

show expanded records and lists inline, so type Tree = { kids: List<Tree> }
recursed until the compiler's stack overflowed. A type met again inside its
own expansion now renders through its one typed render function, which starts
with an empty expansion stack.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ed type [TESTING-EQUALITY]

expect, check and expectError render expected and actual at their own type,
so a failed payload reads 'expected Error(DivZero), got Error(Overflow { by: 3 })'.
An operand with no rendering (a function, a Map) is rejected at compile time.
The printable checks move to builtin_constraints/printable.rs (file < 500 LOC).
…UILTIN-PRINT-FORM]

{ why: "a\"b" } was ambiguous: quotes and backslashes inside a record,
variant or list printed raw. osp_string_show quotes and escapes on GC, ARC
and wasm32. Map stays unprintable; the spec says why (unspecified order).
@MelbourneDeveloper
MelbourneDeveloper merged commit bbfb80d into main Nov 20, 2026
132 of 138 checks passed
@MelbourneDeveloper
MelbourneDeveloper deleted the chore/lspkit-0.10 branch November 20, 2026 09:20
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant