Checker & hover bug fixes, results-first profiler UX, shipped profiling UI, and a truthful profiler docs rewrite#298
Merged
Conversation
collect_unconditional_assigns had no Stmt::Try arm, so a name assigned in both the try body and every except handler was never counted as unconditionally bound, and names_unbound flagged `return o` even though `o` is bound on every path reaching the return. A name is now guaranteed after a try statement when every fall-through path binds it: the success path (try body + else) intersected with each handler, plus anything finally binds. Without handlers an exception propagates out of the function, so the success path alone counts. Closes #285
has_parameterized_bound/has_parameterized_constraint were computed by expr_is_parameterized, which flagged ANY subscripted type outside a typing-forms whitelist — answering "is this generic?" rather than "is this parameterized by a type variable?". PEP 484 permits concrete generic bounds such as Callable[..., Any] or dict[str, int]; only type arguments referencing a type variable (e.g. list[T]) are invalid. The flags are now computed by expr_parameterized_by_typevar, which recurses into the subscript's type arguments looking for an actual type variable reference: a TypeVar declared earlier in the module, or a name matching the single-letter uppercase convention. The obsolete helper and its TYPING_FORMS whitelist are removed. Closes #283
A legacy implicit alias with a structural container body, e.g.
`RefsDictT = dict[tuple[str, str], str]`, was collected by no alias map:
collect_union_aliases kept only Union-bodied definitions, and the
generic-alias map only takes TypeVar-parameterised bodies. The
annotation therefore stayed an opaque Named type, and the empty dict
literal (dict[Never, Never]) was reported as not assignable to it.
The collector (renamed collect_value_aliases) now also accepts concrete
dict/list/set/tuple definitions, excluding TypeVar-parameterised bodies
which remain the generic-alias map's job. The alias expands to its
structural type, where Never vacuously matches any target, so `{}` is
assignable — while genuinely mismatched literals (e.g. an int key
against tuple[str, str]) still fire after expansion.
Closes #282
The blank and caret-underline gutter rows padded the pipe two columns further right than the source row, so every underline rendered shifted two columns off the flagged code (issue #279). All rows now use the rustc-standard width+1 gutter.
A base recorded as an attribute of an unresolved module (class Client(httpx.Client)) resolves by simple name back to the class itself, making the class its own ancestor. Twelve rule/resolver walks recursed on that chain without a cycle guard and overflowed the stack, aborting the LSP in a restart loop on large workspaces such as Apache Airflow (issue #278 — the earlier 64 MiB analysis stacks cannot save an infinite recursion). Predicate-shaped walks now share two cycle-safe helpers in rules/shared.rs (class_or_base_matches, any_base_name_matches); base names pass through verbatim so each call site keeps its exact normalisation and nothing changes but termination. Accumulator and first-match walks in constructors_call_init, dataclasses_transform_class and the resolver's protocol_ext carry inline visited sets. Regression test runs the checker on the minimal airflow shape and also pins the follow-on false positive (a class counting as overriding its own methods through the name collision).
…_ hint Three hover gaps, one shared cause — member information never survived to the hover handler: - issue #287: `Model.model_validate(...)` on a pydantic BaseModel subclass showed nothing. Class methods were dropped at every layer: ExternalSymbol had no member field, stub/py.typed export extraction discarded them, and runtime-generated stubs emitted bodyless `class X: ...`. All three layers now carry methods, and hover resolves dot-access through local base chains (cycle-guarded) into imported classes. - issue #288: `" ".join(...)` showed nothing. New curated typeshed mirror `basilisk_stubs::builtin_method_signature` covers all public str methods; hover types string-literal receivers (with a bytes-prefix guard) and str-annotated/inferred variables. - issue #289: class hovers now append the constructor signature — the class's own `__init__`, else the nearest inherited one in the local base chain. The three lookups share one breadth-first, cycle-guarded base walker. hover.rs (834 lines) is split into hover/{mod,members,tests}.rs to meet the 500-LOC rule.
Stopping a CPU profile previously opened the raw .cpuprofile in VS Code's built-in viewer — a wall-of-numbers self/total-time table — while the Basilisk Profiler results panel hid behind a dismissable toast action. Now the self-contained results panel (summary cards, flame graph hero, click-to-source hot tables) opens automatically on stop, and the native trace stays one deliberate click away: the completion toast's "Open Trace in VS Code Viewer" action, the panel's button of the same name, or the new "Basilisk: Show Profile Results" palette command that re-opens the last result. Snapshot toasts carry a "View Results" action. An explicit native-viewer request that cannot open is said out loud and reveals the trace file instead of failing silently. Also resolves the [PROFILE-UI-GATE] spec/code contradiction: profiling now ships enabled by decision, not by an accidentally-committed override. profiling-ui.ts is deleted, every basilisk.profilingEnabled manifest reference is removed (session-state gating such as basilisk.memoryTracking is kept), the spec section is rewritten to document the shipped state, and a manifest sweep test fails if the dead context key ever returns — a when clause on a key nobody sets would silently hide the UI from every shipped user. The branch-audit conform-5 item is recorded as signed off. Verified: full VS Code extension suite 505 passing; the real CPU e2e asserts the panel opens with zero clicks on an actual debug-launched session and stays reachable after closing.
The page had drifted badly from the implementation and described several features that never existed. Audited every claim against the extension sources, the LSP crates, and LSP-PROFILING-SPEC.md, then rewrote the English page and its Chinese mirror. Removed fabrications: the [tool.basilisk.profiler] pyproject section (profiler config is editor settings only), the CPU "Compare Profiles" command, BSK-PROF-GIL detection, percentage-based leak confidence scores, the basilisk-profiles/ output directory, and the Linux "child shim". Corrected facts: heat thresholds are 20/10/5/1 (not 10/5/2/1); profiler diagnostics are Hints and memory codes are BSK-MEM-* with per-code severities (BSK-PROF-MEM never existed); real VS Code command titles; Zed slash commands are informational and /memsnap //memdiff do not exist; Neovim commands are :BasiliskProfile/:BasiliskMemLeak/ :BasiliskMemStop/:BasiliskMemRefs; artifacts go to the OS temp dir; snapshots write no files; memory profiling requires a debug session. Added the current UX: the Python Processes panel as the primary entry point, the results panel opening on stop with Show Profile Results re-entry, native .cpuprofile/.heapprofile viewers, the cooperative macOS sampler, memory autopilot, final snapshot on exit, the too-short-to-sample message, and the full 12-setting configuration table. The competitor table is reduced to sourced, linked claims per the documentation-honesty rule.
"Open in Speedscope" previously revealed the trace file and dropped the user on speedscope.app's empty Browse page for a manual drag-in, because an https page can never read file:// URLs. Browsers do treat loopback as a potentially-trustworthy origin, so a new in-extension HTTP server (profile-server.ts) serves each exported profile over http://127.0.0.1:<port>/<token>/<basename> and the speedscope deep link loads it automatically. Containment: loopback-only bind, unguessable 32-hex token per file, GET only, no-store, 10-minute registration expiry, torn down with the extension. A companion toast keeps the reveal-and-drag path for browsers without the loopback mixed-content exemption, and a serving failure falls back to it directly. Memory profiling gets the same landing inversion as CPU: taking a snapshot now opens the Basilisk memory dashboard instead of the raw .heapprofile table (presentSnapshot's option renamed to openResultsView). The dashboard offers the raw trace on demand via two new buttons — "Open Heap Profile in VS Code Viewer" and "Open in Speedscope (external)" (speedscope imports V8 sampling heap profiles) — both rendered only when a .heapprofile was written, both routed through the shared openNativeTraceViewer / openSpeedscopeImport helpers. URLs carry the real basename so speedscope's extension-based format detection distinguishes heap profiles from speedscope JSON. Spec sections [PROFILE-VIEWER-DELIVERY], [PROFILE-NATIVE], and [PROFILE-MEMORY-FINAL] rewritten to match; website profiler docs (en + zh) updated. New profile-server.test.ts pins the served headers, token access control, expiry, and teardown; the dashboard-button and webview suites extended; full extension suite green.
The quick-start's profiling step (en + zh) described a process picker that never existed and the old flow: it now walks the Python Processes panel launches, the auto-opening results panel, the real Stop keybinding chord, the debug-session requirement for memory tracking, and the real "Compare Memory Snapshots" command title; the "profile diffing" promise is gone. The Zed install page's slash-command table oversold the profiling/memory commands as actions — they are guides that explain the matching basilisk.profiler.* / basilisk.memory.* LSP commands, and the table now says so. The docs index (en + zh) blurb no longer attributes all CPU profiling to py-spy, since debug-launched macOS sessions use the cooperative in-process sampler.
cargo fmt --all --check flagged seven files with formatting drift from recent work; applied. Coverage ratchets move up to measured minus the 1% buffer: basilisk-cli 89 -> 90 (measured 91), vsix 91 -> 92 (measured 93). Full local CI checklist green at these levels: lint (clippy + eslint + deslop), Zed (wasm build, clippy, 97 tests, mirror standalone build), website (rules.json sync, conformance refs, build, 24 Playwright tests), Rust tests + conformance gate + all crate thresholds, Neovim (43% >= 42%), VS Code e2e suite (93% >= 91%).
…merge Post-merge coverage measured basilisk-cli at 95% and basilisk-stubs at 93%; thresholds move to measured minus the 1% buffer (94 and 92). Full local CI checklist re-run green on the merged tree: format, lint (deslop 12.53% < 12.6% with the current CLI), Zed (all four steps), website (build + 24 e2e), Rust tests + conformance gate + all thresholds, VS Code (510 passing, 93% >= 92%), Neovim (43% >= 42%), Shipwright binary gate, release build, and mutation testing (125 mutants, 118 caught, 0 missed, 100% kill rate, baseline unchanged).
This was
linked to
issues
Jul 8, 2026
This was
unlinked from
issues
Jul 8, 2026
Merged
1 task
abdushakoor12
added a commit
that referenced
this pull request
Jul 9, 2026
…ber hover (#278, #287) (#299) ## TLDR Fixes the LSP stack-overflow restart loop on self-named external bases (#278) and makes external-member hover actually work — pydantic's `model_validate` now hovers on real packages (#287). ## What Was Added? - `chase_py_typed_reexport` / `sibling_module_file` in `crates/basilisk-checker/src/exports.rs`: when a `from pkg import Name` lookup misses in a py.typed package's resolved file, the population pass now follows the `__init__`'s module-level named **and star** re-export imports (including inside `if TYPE_CHECKING:` blocks) into sibling module files and takes the symbol — with its methods — from there. Cycle-guarded via a visited set; runs only on a miss, only for py.typed sources. - Regression tests: - `self_named_external_base_constructor_call_does_not_overflow` (checker) — the `class Client(httpx.Client)` + `Client()` shape that SIGABRT'd the process. - `test_hover_on_method_reexported_through_py_typed_package_init` (hover) — an on-disk pydantic-v2-shaped package: `__init__.py` containing only `if TYPE_CHECKING: from .main import *`, `main.py` defining the class with a `@classmethod`. - `default_mode_analysis_populates_external_imported_symbols` (salsa engine) — default-mode analysis must carry external `imported_symbols`. - `benchmarks/status/darwin-arm64-apple-m4.csv` — first bench baseline for this machine, per the trend policy. ## What Was Changed or Deleted? - **`constructors_call_init` cycle guards (#278):** `find_init_in_hierarchy`, `has_custom_init_in_bases`, and `is_subclass` walked base-class chains with no cycle guard. A class named after its unresolved external base (`class Client(httpx.Client)`) resolves by simple name back to itself in the class map, so constructing it recursed to a stack overflow — SIGABRT, LSP restart loop, unusable editor. All three walks now thread `visited` sets, the same pattern as the twelve walks guarded in #298. Behaviour is unchanged except termination. - **Salsa engine wiring (#287):** `SalsaAnalysisEngine::analyse` populated `imported_symbols` only in `crossModule` analysis mode — which is documented "reserved for future use" and never the default, so the external-member hover shipped in #298 was dead in every real editor session (its unit test injected the symbols by hand). The resolved view handed to hover/completion/navigation now always comes from the memoized `cross_resolved_module` query; **diagnostics remain mode-gated exactly as before**, so non-cross modes keep byte-for-byte CLI parity on what they report. Doc comments updated to match. - Mutation baseline re-dated by the fresh run (identical 100% score, same 125-mutant pool). ## How Do The Automated Tests Prove It Works? - `self_named_external_base_constructor_call_does_not_overflow` aborted the test process with `stack overflow / SIGABRT` before the fix and now passes, pinning both termination and the no-false-positive behaviour (`constructors_call_init` must not fire). - `test_hover_on_method_reexported_through_py_typed_package_init` failed at `imported_symbols.contains_key("BaseModel")` before the extraction fix (and again when the fixture switched to the real star-import shape before the star chase existed); it now asserts hover renders `BaseModel.model_validate` from the re-exported class. - `default_mode_analysis_populates_external_imported_symbols` failed with empty `imported_symbols` before the engine change; it now proves default-mode analysis populates them while the diagnostics path is untouched. - Verified end-to-end over raw LSP against real pydantic 2.13.4: hover on `model_validate` returns `(method) def BaseModel.model_validate(cls, obj: Any, …) -> Self`; `basilisk check` on the crashing repro file went from exit 134 (SIGABRT) to a clean run. - Full local CI pass: lint (fmt/clippy/eslint/deslop), workspace tests + coverage thresholds (all 10 crates), conformance **141/141 = 100%** against freshly fetched `python/typing@main` fixtures, VS Code E2E (510 passing against the staged release VSIX), Neovim (43% ≥ 42%), Zed (97 tests), Shipwright version contracts, mutation testing (125 mutants, 118 caught, 0 missed, 100% kill rate, baseline unchanged), `make bench` PASS, release build. ## Spec / Doc Changes None — code-level doc comments only (`salsa_engine.rs` module/method docs now describe the always-cross resolved view; the new exports helpers document the re-export chase). ## 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-reported bug fixes (four checker/CLI precision bugs, three hover gaps), a profiler UX overhaul that lands every profile on Basilisk's own results views with speedscope loading automatically, the profiling UI officially shipped (availability gate deleted), and a ground-truth rewrite of the profiler documentation — with coverage ratchets moved up to match.
References: issue #285, issue #283, issue #282, issue #279, issue #278, issue #287, issue #288, issue #289 .
What Was Added?
crates/basilisk-lsp/src/hover/members.rs): hovering a method inherited from an external base (e.g.pydantic.BaseModel.model_validate, issue pydantic BaseModel model_validate don't have hover #287) resolves the receiver through a cycle-guarded, BFS base-class walk intoimported_symbols; builtin receivers get signatures from a new curated table of all 47 publicstrmethods (crates/basilisk-stubs/src/builtin_members.rs, issue method don't have on hover #288); hovering a class name now appends its own-or-inherited__init__signature (issue hover on class don't have init hint #289).ExternalSymbolgains amethods: Vec<ExternalMethod>field, populated by workspace export extraction,.pyiparsing, and runtime-introspection stub generation (which now emits class bodies instead ofclass X: ...).vscode-extension/src/profile-server.ts): serves exported profiles overhttp://127.0.0.1:<port>/<token>/<basename>so the speedscope.app deep link loads automatically instead of dumping the user on the empty Browse page. Loopback-only bind, unguessable 32-hex token per file, GET-only,no-store, 10-minute expiry, torn down on deactivate.crates/basilisk-checker/src/rules/shared.rs::class_or_base_matches): one transitive, cycle-guarded walk now backs guards, missing-override, TypedDict-required, NamedTuple, protocol, enum, and dataclass-transform rules that each had a private copy.What Was Changed or Deleted?
try/except/else/finallyare computed path-correctly, endingnames_unboundfalse positives on try-assigned names (issue False positive "returnsobutomay be unbound on some paths" #285);TypeVar(bound=Callable[..., Any])parses instead of trippinggenerics_basic(issue False positive "has abound=that is parameterized by a type variable" #283); an emptydict()/{}is assignable to adicttype alias (issue Incorrect assignability with dict alias #282); CLI snippet gutters align with the source row (issue Red squiggles are misaligned #279); base-walks stay cycle-safe when a class shadows its external base's name (issue Stack overflow on VS Code when opening a large project #278 discipline)..cpuprofiletable; taking a memory snapshot opens the Basilisk memory dashboard instead of the raw.heapprofile. The native V8 viewers remain one deliberate click away everywhere, and an explicit request that cannot open is announced and reveals the file — never a silent failure.profiling-ui.ts(the test-only availability gate, left force-enabled with a "do not commit" comment) is deleted; all 24basilisk.profilingEnabledmanifest references removed, keeping only real session-statewhenclauses. The spec section now guards the residue risk: awhenclause on a context key nobody sets silently hides UI from every shipped user.hover.rssplit intohover/{mod,members,tests}.rsto respect the 500-LOC file limit.How Do The Automated Tests Prove It Works?
names_unbound_tests(try/except binding, issue False positive "returnsobutomay be unbound on some paths" #285),generics_basic_tests(TypeVarwithCallablebound, issue False positive "has abound=that is parameterized by a type variable" #283),assignment_compatibility_tests(empty-dict → alias, issue Incorrect assignability with dict alias #282), and the CLI gutter-alignment assertion inbasilisk-cli/src/output/mod.rs(issue Red squiggles are misaligned #279).test_hover_on_method_inherited_from_external_stub_base_shows_signaturebuilds a realpydantic.pyi, resolves imports, and asserts the fullBaseModel.model_validate(obj: object) -> BaseModelsignature;test_hover_on_str_literal_method_shows_signatureassertsstr.joinon a string literal;test_hover_on_class_shows_init_signatureasserts the__init__block;entries_to_pyi_emits_class_methodspins runtime-stub method emission.basilisk.profileShowResultsafter closing;profile-server.test.tspins the served body, CORS header,no-store, 404-on-wrong-token, registration expiry, and post-dispose unreachability; webview unit tests pin every new button (and their absence when no trace exists); a manifest sweep test fails ifbasilisk.profilingEnabledever reappears.python/typingscorer; mutation testing 125 mutants — 118 caught, 0 missed, 100% kill rate (baseline unchanged); VS Code suite 510 passing at 93% coverage; Neovim 43%; all Rust crates at or above their (raised) thresholds; deslop duplication gate at 12.53% < 12.6%.Spec / Doc Changes
docs/specs/LSP-PROFILING-SPEC.md:[PROFILE-UI-GATE]rewritten as shipped-enabled UI availability;[PROFILE-NATIVE]/[PROFILE-NATIVE-FALLBACK]describe the results-view-first routing;[PROFILE-VIEWER-DELIVERY]documents the loopback deep-link flow and its containment;[PROFILE-MEMORY-FINAL]matches the dashboard-first landing.docs/PROFILER-BRANCH-AUDIT.mdrecords the conform-5 sign-off.website/src/docs/profiler.md(+ zh mirror) rewritten against the implementation: removed fabricated content (pyproject profiler config, CPU "Compare Profiles", GIL detection, percentage leak scores,basilisk-profiles/directory), corrected heat thresholds (20/10/5/1), diagnostic codes/severities (BSK-PROF-*Hints,BSK-MEM-*family), command names across VS Code/Zed/Neovim, and added the Python Processes panel flow, cooperative macOS sampler, memory autopilot, and the full 12-setting configuration table.quick-start.md,install-zed.md, and the docs index (en + zh) swept for the same stale claims.Breaking Changes