Six bug fixes: hover inference, glyph tallies, safe-only Fix All, uv import isolation, Modules loading state, deep-expression stack overflow#294
Merged
Conversation
Hover on an unannotated function previously rendered a bare, type-less signature. Now: - The return type is inferred from the body's return statements and rendered when the annotation is missing (shared with inlay hints via the hoisted util::infer_return_type_display). - Unannotated parameters render ': Unknown' instead of blank, except the implicit self/cls receiver. - classify_rhs resolves two-argument <dict-literal>.get(key, default) calls to a new RhsKind::KnownCall carrying the common value/default kind, so the issue's repro infers '-> str'. The checker maps KnownCall exactly like CallExpr — display-only, no diagnostic or conformance change. Adds a stdio LSP e2e regression test on the exact repro and documents the inferred-display behavior in [LSPARCH-FEATURES-HOVER].
…LES-COUNT-STYLE] (#236) diagnosticTally emitted the forbidden lettered `nE nW` form on all three Modules-panel surfaces (module rows, folder/package rollups, workspace header message). Switch it to the spec-mandated coloured glyph tally, omitting zero severities. Adds a regression e2e test, flips the five existing assertions that encoded the old format, restores the regressed [EXTACT-MODULES-TREE-STRUCTURE] spec bullet, and regenerates vscode-module-explorer.png via the sanctioned screenshot pipeline.
…ASSIFY] (#245) All three LSP fix-all surfaces (source.fixAll code action, basilisk.fixFile, basilisk.fixWorkspace) applied the unfiltered ALL_FIXABLE_RULES set, silently including the Unsafe BSK-E0003 `: Any` insertion, and the Safe/All command split promised by [AUTOFIX-MASS-VSCODE] did not exist. Route the default surfaces through fix_filtered_in_file(SAFE_FIXABLE_RULES) — the machinery the CLI already uses — and add explicit all-tier basilisk.fixFileAll / basilisk.fixWorkspaceAll commands (constants, dispatch, VS Code palette entries, URI-injection middleware). Adds a stdio e2e regression test covering all five surfaces, updates the [AUTOFIX-CONFLICTS] honest-state note and the retitled Fix All (Safe) references in the activity-panel spec and README.
…nterpreter (#252) The lockfile E0010 tests were not hermetic: a uv-locked project with no venv fell back to `python3` sys.path discovery, so `import flask` resolved on any machine whose first ambient site-packages carried flask and `make test` went red (green machines passed only by sys.path ordering luck). Gate the ambient-interpreter fallback off when a PackageRegistry exists — the lock is the source of truth, so a dependency missing from it now diagnoses identically on every machine. In-tree venv discovery and an explicitly activated VIRTUAL_ENV (issue #25) still apply to locked projects. States the lock-only resolution contract in [LSPUV-DIAGNOSTICS-MODULE-NOT-FOUND] and adds a deterministic regression test at the search_paths_from_config seam.
…the scan finishes (#144) The Modules panel rendered the terminal 'No Python files found' empty-state whenever the workspace rollup reported totalFiles == 0 — including the whole window between the server reaching Running and its background workspace scan finishing, and it showed nothing at all while the client was still starting. Server: LspServer tracks initial_scan_complete; every basilisk.workspaceModules rollup stamps scanComplete, and a new basilisk/scanComplete notification fires when a scan finishes (set immediately in openFilesOnly, where no scan runs). The notification is required, not an optimisation: a genuinely empty workspace publishes no diagnostics, so nothing else would settle the loading state. Client: workspaceHealthMessage renders 'Analyzing workspace…' when no stats exist yet or totalFiles == 0 without scanComplete; the #57 empty-state now requires a finished scan. The store bumps analysisRevision on scanComplete, so the repaint flows through the existing signals effect — no polling, no panel-local readiness flags. Spec: [EXTACT-MODULES-HEADER-LOADING] loading clause, [EXTACT-LSP-COMMANDS-SCAN-COMPLETE] notification, HealthStats.scanComplete in both wire-shape sections; e2e regression tests drive the real provider through the tree-view chrome for the mid-scan, pre-ready, and genuinely-empty states.
…shes analysis (#278) A single generated file like `total = 1 + 1 + …` parses flat but builds an arbitrarily deep left-nested BinOp tree; the recursive resolver/checker visitors then overflowed the default ~2 MiB tokio worker stack during the workspace scan, aborting the whole LSP server and crash-looping the editor's restart logic (0xC00000FD / 'tokio-rt-worker has overflowed its stack'). The CLI had the same exposure on the process main thread. Two complementary mechanisms per [LSPARCH-ARCH-STACK] and the extended [CHKARCH-ARCH-PARSEDEPTH]: - Every thread that can run analysis now has a 64 MiB stack: the new basilisk-lsp runtime module hosts block_on and all tokio workers on analysis-sized stacks for both LSP entry points, and the CLI dispatches every subcommand through its synchronous counterpart. - The parser's linear pre-parse nesting guard (previously brackets and indentation only — its spec listed this construct as a known residual) now also caps depth-building operator chains at 50,000 per expression context, rejecting pathological files with 'expression too deeply nested' before any recursion can start. Flat constructs (and/or, comparisons, comma lists) are exempt so giant generated literals stay analysable. Limits sit >2x under the measured 64 MiB crash floor (~150k). Tested with real-binary e2e coverage: the LSP scan survives and stays responsive on a 10,000-term chain, basilisk check analyses 30,000 terms cleanly (previously SIGABRT), 300,000 terms is skipped with a warning like the bracket cap (previously SIGABRT), plus operator-chain boundary tests in the parser suite. The stdio LSP test harness moved to a shared module. Bench gate passes with no regression; conformance holds at 100% (141/141), FP 0.
… [PROFILE-HELPER-PROTOCOL-ERRORS] A helper exiting between the attach payload and newline writes hits the sender as EPIPE, not the reader as EOF; the send path bailed with the bare IO error instead of reaping the helper and surfacing its exit status and stderr (#81 contract, raced on slow CI runners). Also joins four lines that rustfmt 1.96 (CI stable) formats differently than 1.95.
MelbourneDeveloper
added a commit
that referenced
this pull request
Jul 7, 2026
The main merge (PR #294) brought code below three crates' coverage thresholds. Add behavioral tests to restore them; thresholds unchanged: - basilisk-parser: unit tests for depth::check_nesting (bracket/indent/ operator-chain limits + boundaries, dedent, bracket-resume, chain-breakers) — depth.rs 45% -> 100%. - basilisk-resolver: function-body TypedDict tests exercising the recursive check_td_stmts (del of required/NotRequired keys, reads, subscript-assign, disallowed mutator methods) — 94% -> 95%. - basilisk-cli: stub-subcommand tests (find_package_source success, cache_stub, run_stubs* paths) — 87% -> 91%. Authoritative scripts/test-rust.sh: all coverage thresholds pass, conformance 141/141 (100%, 0 false positives).
1 task
MelbourneDeveloper
added a commit
that referenced
this pull request
Jul 7, 2026
…esh (+ main merge) (#295) ## TLDR Consolidation of the `cleanup` branch — expands the benchmark-fixture suite, splits `FsCache` into its own module (500-LOC cap), refreshes the READMEs/website, ratchets the deslop duplication gate to 12.6% and simplifies its installer — with current `main` merged in and every per-crate coverage threshold restored. ## What Was Added? - **12 benchmark fixtures** under `benchmarks/fixtures/*.py` (~24k lines) expanding the perf suite (aliases, callables, constructors, dataclasses, enums, generics defaults, literals, narrowing, overloads, protocols, returns, tuples). - **`crates/basilisk-checker/src/imports/fs_cache.rs`** — `FsCache` extracted into its own module so `resolve.rs` stays under the 500-LOC cap; its tests move with it. - **Coverage tests** restoring three crates that the `main` merge dropped below threshold: - `basilisk-parser`: unit tests for `depth::check_nesting` (bracket/indent/operator-chain limits + boundaries, dedent, bracket-resume, chain-breakers). - `basilisk-resolver`: function-body `TypedDict` tests exercising the recursive `check_td_stmts` (required/`NotRequired` key deletes, reads, subscript-assign, disallowed mutator methods). - `basilisk-cli`: stub-subcommand tests (`find_package_source`, `cache_stub`, `run_stubs*`). - **`examples/cpu_demo_loop.py`** demo. ## What Was Changed or Deleted? - **FsCache split**: `resolve.rs` imports `FsCache` from `fs_cache`; `apply.rs` imports it from `super::fs_cache`. - **Checker rule refinements**: `generics_type_erasure`, `callables_kwargs`, `specialtypes_never_2`, `generics_basic_3/helpers` precision changes. - **Flaky DAP test hardened**: `debug_spawn.rs` `a_missing_interpreter_fails_fast_without_port_retries` uses 3 candidate ports so the documented free-port TOCTOU can't surface `PortTaken` before the spawn (assertion unchanged). - **Deslop gate ratcheted**: `.deslop.toml` `max_duplication_percent` → 12.6 with non-product excludes; `scripts/install-deslop.sh` reduced to an unpinned `brew install nimblesite/tap/deslop`. - **README refresh**: concise `vscode-extension/README.md`; root `README`/`README.zh`; website conformance/benchmark data. - **Merge of `main`** (PR #294: hover inference, glyph tallies, safe-only Fix All, uv import isolation, Modules loading state, deep-expression stack-overflow guard). ## How Do The Automated Tests Prove It Works? - **Conformance** (`conformance/score.py --gate`): **141/141 (100%), 0 false positives, 0 missed required errors** against `python/typing@main` (`f4f2952`). - **Per-crate coverage** (`scripts/test-rust.sh`, thresholds in `coverage-thresholds.json` **unchanged**): all pass — `basilisk-parser` 100%, `basilisk-resolver` 95%, `basilisk-cli` 91%, `basilisk-checker` 94%, plus db/lsp/mojo/parser/plugin/stubs/config. - **Lint**: `cargo fmt --all --check`, `cargo clippy --workspace --all-targets -- -D warnings`, and the deslop gate (repo-wide duplication under the 12.6% ceiling) all clean. - **Flaky fix**: the `debug_spawn` suite passes across repeated runs that previously flaked under port contention. - **Website**: `gen_conformance_reference.py --check` and `gen_rules_reference.py --data` drift checks clean; 24 Playwright navigation/screenshot e2e tests pass. ## Spec / Doc Changes - `docs/specs/CHECKER-ARCHITECTURE-SPEC.md` conformance stamp updated to the graded commit. - READMEs (`README.md`, `README.zh.md`, `vscode-extension/README.md`) and website docs refreshed. - `.deslop.toml` duplication budget documented (ratchets down only). ## Breaking Changes - [x] None
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
TLDR
Six user-facing bug fixes across the LSP, CLI, parser, and VS Code extension: hover type inference for unannotated functions, glyph-styled module tallies, safe-only Fix All defaults, uv-locked import isolation, an honest Modules-panel loading state, and a two-layer defence against the deep-expression stack overflow that crash-looped the server.
Closes #253
Closes #236
Closes #245
Closes #252
Closes #144
Closes #278
What Was Added?
crates/basilisk-lsp/src/runtime.rs(new, implements[LSPARCH-ARCH-STACK]): every production entry point now runs analysis on 64 MiB stacks —run_with_analysis_stackhosts the CLI's work thread, andblock_on_with_analysis_stackbuilds the tokio runtime (block-on thread + workers both sized) for the stdio and WebSocket LSP servers. BareRuntime::new()is gone from production paths (Stack overflow on VS Code when opening a large project #278).crates/basilisk-parser/src/depth.rs(extends[CHKARCH-ARCH-PARSEDEPTH]): the linear pre-parse lexer scan now counts depth-building tokens (binary/unary operators,., ternary,lambda) per bracket level and rejects chains past 50,000 with "expression too deeply nested" — the construct the spec previously documented as a known residual. Flat-by-construction operators (and/or, comparisons, commas) are exempt so giant generated literals still analyse.RhsKind::KnownCallin the resolver: two-argument<dict-literal>.get(key, default)calls resolve to their common value/default type, so hover infers-> strfor the Hover shows no inferred types for unannotated function params/return (def f(missed, caught, ...)) #253 repro. Display-only — no diagnostic or conformance change.[AUTOFIX-CLASSIFY]): four commands (fixAllSafe/fixAll× file/workspace) inbasilisk-common, the LSP command handlers, and the VS Code palette.source.fixAlland the default palette action apply onlySAFE_FIXABLE_RULES; BSK-E0003's: Anyfix is unsafe-tier only ([AUTOFIX-CLASSIFY] LSP fix-all surfaces apply the Unsafe BSK-E0003 fix by default; promised Safe command variants don't exist #245).scanCompleteflag on the LSP's module-stats payload plus abasilisk/scanCompletenotification, consumed by the extension store (Modules panel falsely shows "No Python files found" while the analyzer is still starting up #144).crates/basilisk-cli/tests/lsp_stdio/mod.rs, hoisted frome2e_lsp_no_ruff.rsand reused by the new deep-expression suite.What Was Changed or Deleted?
basilisk-lsp/src/util.rs, shared via hoistedinfer_return_type_display): unannotated parameters render: Unknown(except the implicitself/clsreceiver) and missing return annotations render the type inferred from the body'sreturnstatements instead of a bare, type-less signature (Hover shows no inferred types for unannotated function params/return (def f(missed, caught, ...)) #253).basilisk-lsp/src/import_resolver.rs,[LSPUV-DIAGNOSTICS-MODULE-NOT-FOUND]): a project with auv.lockresolves third-party imports only from its own environment — the ambient-interpreter site-packages fallback is disabled, so undeclared dependencies are correctly flagged instead of silently resolving from the global interpreter (Lockfile E0010 tests are not hermetic: fail on any machine with flask installed globally (blocks make test) #252).vscode-extension/src/module-explorer.ts,[EXTACT-MODULES-COUNT-STYLE]): rows render coloured glyphs🔴 n/🟠 n, omitting zero severities, instead ofnE nWtext ([EXTACT-MODULES-COUNT-STYLE] module tallies rendernE nWtext instead of required glyph tally #236).[EXTACT-MODULES-HEADER]): "No Python files found" is only shown after the workspace scan reportsscanComplete; during the scan the panel says "Analyzing workspace…" (Modules panel falsely shows "No Python files found" while the analyzer is still starting up #144).run_commandon the analysis-stack thread; LSP stdio/WS entry points route through the shared runtime module (Stack overflow on VS Code when opening a large project #278).How Do The Automated Tests Prove It Works?
e2e_deep_expressions.rs(real binary over stdio):workspace_scan_survives_deeply_chained_binary_expressionopens a workspace whose file is a 10,000-term1 + 1 + …chain — the pre-fix server aborts withstack overflow; the test requiresbasilisk/scanCompleteplus a hover response.cli_check_survives_deeply_chained_binary_expressionchecks 30,000 terms exits 0, andcli_check_rejects_pathological_expression_depth_instead_of_crashingrequires 300,000 terms to be skipped with "expression too deeply nested" rather than crash.parse_tests.rspins the guard boundary: 50,000 operators parse, 50,001 report the message, comma-separated flat sums and bracket-isolated sub-chains stay accepted, unbalanced closers don't disarm the counter, and 60,000-deep unary/attribute/ternary/lambda chains are all rejected.lsp_e2e_hover.rs::test_lsp_hover_unannotated_function_shows_inferred_typesasserts the Hover shows no inferred types for unannotated function params/return (def f(missed, caught, ...)) #253 repro hovers as(function) def extracted_function(missed: Unknown, …) -> str.lsp_e2e_code_actions.rsasserts thesource.fixAllsurface excludes the unsafe BSK-E0003 edit while the explicit all-fixes command includes it.module-explorer-loading.test.tswalks the Modules panel falsely shows "No Python files found" while the analyzer is still starting up #144 lifecycle: pre-scan state renders "Analyzing workspace…", post-scanCompleteempty workspace renders "No Python files found";module-explorer-tree.test.tsandactivity-panel.test.tsassert the 🔴/🟠 glyph format and zero-severity omission.-D warnings.Spec / Doc Changes
LSP-ARCHITECTURE-SPEC.md: new[LSPARCH-ARCH-STACK]section — analysis stack sizing, the entry-point contract (never bareRuntime::new()), cross-reference to the parser guard.CHECKER-ARCHITECTURE-SPEC.md:[CHKARCH-ARCH-PARSEDEPTH]gains the operator-chain limit (replacing the "known residual" note) with counted tokens, reset tokens, and exemption rationale.LSP-MASS-AUTOFIX-SPEC.md([AUTOFIX-CLASSIFY]): Safe vs Unsafe tier definitions and the four-command surface.LSP-UV-INTEGRATION-SPEC.md([LSPUV-DIAGNOSTICS-MODULE-NOT-FOUND]): uv-locked resolution contract.EXTENSION-ACTIVITY-PANEL-SPEC.md:[EXTACT-MODULES-COUNT-STYLE]glyph format and[EXTACT-MODULES-HEADER]scan-gated empty state.vscode-extension/README.md+ regeneratedvscode-module-explorer.pngscreenshot reflecting the glyph tallies.Breaking Changes
basilisk.fixFile/basilisk.fixWorkspacecommands keep their IDs but now apply safe fixes only (the intended default per[AUTOFIX-CLASSIFY]); the previous unsafe-inclusive behaviour remains available via the new explicitbasilisk.fixFileAll/basilisk.fixWorkspaceAllcommands.