wolf-lsp

The wolf editor layer: clients, configs, conformance testing, and packaging for the wolf language server.

The language server itself lives in the compiler. wolf lsp is the wolf compiler serving the Language Server Protocol from the same code that compiles, so the editor and the build see the same diagnostics (wolf-lang's wolf_query contract). This repo makes editors talk to it:

• a protocol conformance harness: recorded JSON-RPC session replay against wolf lsp, capability snapshots, latency budgets
• maintained clients for fackr, facsimile, Neovim and VS Code
• configuration for Helix, Zed and Emacs (eglot), and documentation for JetBrains (LSP4IJ)
• marketplace and packaging for all of the above

docs/MATRIX.md lists which editors are supported, at what verification level, and when each was last checked. Every row names its CI job or states that it has never been run.

None of the clients have been published to a marketplace, registry or mirror yet, and there is no tag. The pipelines exist and run offline; each stops at a step that needs a person.

Licensed under GPL-3.0-or-later.

The harness

lspconf drives a real wolf lsp child process over stdio. Half of it needs no server and runs anywhere. The other half needs a binary at the pin and, without one, exits 77 with the reason.

lspconf verify              transcripts and scripts: parseable, valid, canonical
lspconf profiles            capability profiles, and which clients are unread
lspconf doctor              the pin, the binary that won, the verdict
lspconf record <s.lsps>     drive a scripted session, write <s>.jsonl
lspconf rerecord [dir]      re-record every script beside its transcript
lspconf replay [path…]      drive recorded sessions, match per record
lspconf onetruth [sample…]  publishDiagnostics == conform-run, per sample (D34)
lspconf bench [--out F]     latency budgets, D5 JSONL (report-only)
lspconf fuzz [--seed N]     seeded partial-edit session, three oracles

A transcript (transcripts/**/*.jsonl) is a recorded JSON-RPC session, normalized so that it changes when behaviour changes and stays the same when incidental output changes. Beside each one is the script (.lsps) that produced it, so re-recording is one command and its diff is reviewable. Transcripts are generated, and lspconf verify rejects one that is not in canonical form or has no script beside it.

Running the server lane locally

CI does not build the compiler. wolf-lang is a binary dependency here, and building the whole compiler in every job would cost minutes to produce what the acquisition step downloads (vendor/README.md has the detail). Locally it is three commands, once:

git submodule update --init upstream
(cd upstream && cargo xtask dist)
export WOLF_BIN="$PWD/upstream/target/release/wolf"

The build stamp matters (D57), and it has to come from the builder. An unstamped build prints +dev.unknown, the same string every trunk build of that crate version prints, and doctor rejects it: a version string that cannot name its commit is the stale-binary problem the pin exists to prevent.

Two things about the upstream stamp explain why this is xtask dist and no longer a hand-rolled cargo build with WOLF_COMMIT set:

• A release stamp needs WOLF_RELEASE=v{version} as well as WOLF_COMMIT, and upstream grants it only when that exact tag points at HEAD. When the pin is a release tag, a WOLF_COMMIT-only build prints <version>+dev.<sha>, which does not match the bare version the PIN records.
• The old recipe abbreviated the sha itself, as --short=7. Upstream's stamp uses --short with git's automatic width, which is eight for wolf-lang today, so the two differed by one character and doctor rejected the difference. Running upstream's builder gives upstream's answer.

cargo xtask dist also stages a tarball under upstream/target/dist/. The binary at upstream/target/release/wolf is the same build.

Then cargo run --bin lspconf -- doctor should say READY, and the gated commands work:

cargo run --bin lspconf -- --require-server replay     # the transcript library
cargo run --bin lspconf -- --require-server onetruth   # D34
cargo run --bin lspconf -- --require-server fuzz regions.lu --seed 1
cargo xtask type-names-check --require-server          # TYPE_NAMES against the compiler (#24)
cargo test                                             # the gated suites go live

Without WOLF_BIN, resolution falls back to .wolf-bin/, the artifact cache CI uses, and then to wolf on PATH. doctor reports which one won, because "works on my machine" is usually a second wolf earlier in PATH.

When the build and the editor disagree

lspconf onetruth tests D34 directly. For every sample it runs wolf conform-run --error-format=json and an LSP session over the same bytes, then checks that the diagnostics agree: same codes, same spans (through the negotiated position encoding), same messages, and reachable from some open document of the module.

A mismatch is a wolf-lang bug and is filed upstream with both records attached. This repo does not normalize or patch around it, because detecting divergence is what the harness is for. divergences.toml is the ledger of filed mismatches. An unfiled divergence fails the gate, and so does a ledger entry whose bug has since been fixed.