read_file determined line indices via Python's splitlines, which breaks on
\f, \v, \x1c-\x1e, \x85, U+2028 and U+2029 in addition to the LSP-compliant
line breaks counted by delete_lines/replace_lines. In a file containing any
of those characters the indices reported by read_file referred to different
physical lines than the editing tools operate on, so an edit made after the
required read_file could silently hit the wrong line.
Use TextUtils.split_lines, as already applied to search_text and
MatchedConsecutiveLines.from_file_contents. Files whose line breaks are all
LSP-compliant are decomposed exactly as before.
delete_lines/replace_lines address the position one line past the last line
(line == number of lines, col 0). get_index_from_line_col cannot resolve that
position when the file has no trailing newline (no closing newline to count),
so delete_text_between_positions raised InvalidTextLocationError and the delete
crashed on a very common file shape.
Clamp that past-EOF end position to end-of-file in delete_text_between_positions,
mirroring the guard insert_text_at_position already has. A file with no trailing
newline now deletes its last line identically to one with a trailing newline.
Adds TextUtils unit tests and a CHANGELOG entry.
stepping line by line instead of character by character,
delegating the scan for line breaks to native string search
(~15-100x speedup for a full pass over typical source files)
* Add helper class TextStepper to facilitate computations
* Reimplement TextUtils methods based on TextStepper
* Adhere to LSP definition of newlines being defined by either "\n", "\r\n" or "\r" on its own.
Previously, only "\n" was considered.
* Add util method TextUtils.split_lines, allowing usages of Python's splitlines to be replaced
The cpu matrix job had no timeout-minutes, so a hung job runs until
GitHub's default 6-hour limit. This is not hypothetical: on run
29064865197 (2026-07-10), the native (ubuntu-latest) job hung and ran
from 02:36:34 to 08:36:50 before GitHub cancelled it at exactly the
6-hour cap, blocking the PR's status the whole time and burning a
runner for 6 hours.
Healthy runs are far below the new limit: across the 8 most recent
successful runs of this workflow, the slowest job was 16 minutes
(niche on ubuntu-latest), with catch-all on windows-latest typically
11-14 minutes. 60 minutes leaves ~4x headroom for cache-cold runs
(language-server downloads, npm installs) while cutting the worst-case
hang from 6 hours to 1.
Part of the timeout-hardening discussed in #1695.
When auto-generating a Serena project configuration at startup, compute the languages
in a background thread and update them later to avoid delays of the MCP server startup.
Where completeness of the language list matters (i.e. language server initialisation),
a corresponding event is awaited to ensure completeness.
Perl::LanguageServer only indexes files whose extension is in its
`perl.fileFilter` and skips directories in `perl.ignoreDirs`. Both were
hardcoded in PerlLanguageServer, so projects whose Perl source uses
non-standard extensions (e.g. .cgi / .psgi web handlers, which can
dominate a mature codebase) had no way to make those files visible to
symbol/reference queries without forking Serena (#1449).
Surface both lists through `ls_specific_settings["perl"]` (keys
`file_filter` and `ignore_dirs`), defaulting to the existing values so
behaviour is unchanged for current projects. The resolution lives in a
pure static method `_resolve_filter_settings` so the config plumbing is
covered by unit tests that don't require a Perl runtime.
Keep Serena's symbol index in sync with the LS (#1449, per review):
`find_symbol` and symbol indexing are driven by
`Language.PERL.get_source_fn_matcher()`, so adding an extension (e.g.
.cgi) to `file_filter` alone would leave the matcher untouched and the
extension's symbols invisible. Since that matcher is now a @cache'd
per-language singleton, add a `FilenameMatcher.add_extensions()` mutator
and call it from `PerlLanguageServer` so the configured extensions also
reach the symbol index, the ignore checks, and language composition.
Avoid cross-project leakage (per review): because the matcher is a
process-wide singleton, extensions added for one project would otherwise
persist into the next. `FilenameMatcher` now snapshots its initial
configuration and exposes `reset()`, and `SolidLanguageServer.__init__`
resets the matcher right after `self.language` is determined, so every
activation starts from the language's defaults before the subclass
re-applies its own `file_filter`.
This complements 860c5841 (which added .t to the default list as a
surface-level fix) by addressing the root configurability gap that the
maintainer acknowledged in #1449.
Hanging was never reproduced, this is preemptive hardening
- Add a 120s timeout to the server_ready wait (previously waited with no
timeout, so a server that crashed or never sent experimental/serverStatus
with quiescent=true would hang startup forever). On timeout we log a
warning and proceed, matching kotlin/matlab/powershell/etc.
- Harden the language/status and experimental/serverStatus notification
handlers to use dict.get() instead of raising KeyError on unexpected
payloads (a KeyError in the quiescent handler could itself prevent
server_ready from ever being set, feeding the hang).
- Relax the over-strict completionProvider capability check: an exact-dict
equality assertion crashed initialization whenever rust-analyzer changed
or added a completionProvider field. These values are not used downstream,
so a mismatch is now logged as a warning instead of asserted.
Language servers can report locations of generated files that are not on
disk. JDTLS does this for Lombok-generated types, reporting e.g.
target/classes/test_repo/LombokModel$LombokModelBuilder.class while
indexing; is_ignored_path then raised FileNotFoundError, turning an
intermittent indexing state into a hard test failure
(test_bare_symbol_names[java]; Tests run #4240, job jvm (macos-latest)).
Changes:
- is_ignored_path classifies missing paths with the usual lexical rules
instead of raising: ignored-dirname components, the unsupported-extension
rule, then the configured pathspec. A missing path is treated as a file
if its name has a suffix and as a directory otherwise (so the leaf of a
missing suffixless path is also checked against ignored dirnames).
Behavior for existing paths is unchanged: the checks were reordered
(ignored-dirname before unsupported-extension), but each check is pure,
so the outcomes are identical.
- SymbolLocationRequest.convert_location_item skips locations whose
absolute path does not exist (log + return None), so phantom build
artifacts never reach consumers regardless of ignore configuration.
- New unit tests (test double, no language server process; they run in
the catch-all CI batch) pin the missing-path classification — including
the exact Lombok .class path shape from the CI failure — and the
skipping of symbol locations that are not on disk.
Notes:
- The removed raise arrived verbatim with the multilspy import (2f607ff2)
and was never a deliberate guard.
- The Project layer's equivalent check already tolerates missing paths
(9c19d5f7), but returns False unconditionally. The LS layer instead
classifies lexically so that LS-reported build artifacts still classify
as ignored: with #1645 having removed the JDTLS build-dirname ignores,
it is the unsupported-extension rule that catches *.class paths.
Make the Svelte companion TypeScript server raise on readiness and
indexing timeouts instead of silently proceeding with a partial index,
turning flaky/wrong cross-file results into clear failures:
- TypeScriptLanguageServer: timeout handling factored into overridable
hooks (_handle_server_ready_timeout, _handle_project_indexing_timeout);
the base server keeps its historical permissive behavior, so plain
TypeScript setups are unchanged.
- SvelteTypeScriptServer: both overrides raise TimeoutError; only the
indexing-timeout error includes describe_indexing_state(), instead of
serving requests from a cold or partially indexed program.
- Configurable server_ready_timeout / indexing_timeout via
ls_specific_settings.
- SvelteLanguageServer: file-open failures during companion preparation
raise SvelteCompanionPreparationError with a count-first file listing
(capped at 10), chained to the first underlying error so the cause is
visible without DEBUG logging.
Apply the same fail-loudly principle to the test fixture bootstrap:
- Svelte test repo: npm ci with a committed package-lock.json, removing
dependency-resolution non-determinism from CI runs.
- Run svelte-kit sync explicitly during fixture setup and fail loudly if
.svelte-kit/tsconfig.json cannot be generated. The fixture's "prepare"
script masks sync failures ("svelte-kit sync || echo ''"), and a known
npm bug (npm/cli#4828) can silently skip platform-specific optional
dependencies; the result is unresolvable $lib path aliases and
cross-file tests failing with partial results instead of a clear,
actionable error.
- codespell: skip third-party fixture repositories, using the anchored
pattern ./test/resources/repos (codespell matches skip globs against
the ./-prefixed paths it walks, so an anchor-less pattern never
matches).
Also add direct regression tests for the timeout-policy split
(test/solidlsp/test_typescript_timeout_policy.py: base stays
permissive, companion raises, settings plumbing and precedence, and
start-or-completion indexing-wait semantics). The tests use no mocking or
patching of any kind; a hand-written call-recording fake drives the companion
guard against real file discovery under a pytest tmp_path. They spawn no LS
process and avoid external readiness races; one wait test uses a real Timer
with a 30s ceiling.
The configuration docs state that the TypeScript table lists both knobs; the
Svelte table lists indexing_timeout, with readiness inheritance documented in
prose.
The Nomic solidity-language-server validates asynchronously: document
changes are debounced (~250ms), validation runs a forge/solc compile,
and completion is signalled with the custom notification
custom/validation-job-status -- emitted on every completion path
(success, validation failure, job errors; verified in the bundled 0.8.4
package, which publishes the document's diagnostics before the signal).
Serena ignored the signal ("Unhandled method
'custom/validation-job-status'") and waited a fixed 2.5s for a
non-empty publishDiagnostics, returning [] whenever validation finished
later. Warm environments fit the window; slow or cold ones (macOS and
Windows CI runners, which download solc 0.8.20 on every run -- the CI
cache only covers ~/.serena/language_servers/static, not
~/.local/share/svm) often do not. This is the root cause of the
empty-diagnostics failures of test_file_diagnostics[solidity]: xfailed
on macOS since its introduction (bd7e64e8) and failing on Windows on
Tests run #4236, job catch-all (windows-latest). Shrinking the wait
reproduces the exact failure signature (AssertionError: []) on Linux.
Changes:
- Handle custom/validation-job-status (threading.Event + last payload).
The payload carries no document URI, so the event is global to the
server instance.
- request_text_document_diagnostics keeps the 2.5s non-empty publish
wait as the fast path. When it misses, it awaits the completion
signal against a 60s deadline (aligned with the indexing wait) and
takes the generation-advancing publication for the request URI,
accepting an empty result as the validated answer; a completion that
brings no publication for the URI (e.g. one belonging to another
document) re-arms the wait until the deadline. If the signal never
arrives, a warning is logged and the previous empty-result behavior
applies, bounded by the deadline.
- Remove the darwin xfail from test_file_diagnostics; add regression
tests for the fallback delivery (fast path forced to miss), for the
never-signalled bound, and for re-arming on a spurious completion
belonging to another document.
- CHANGELOG entry under Language Servers.
Design notes:
- Mirrors #1639 (TypeScript/VTS: event-based indexing wait replacing a
fixed sleep).
- The simpler alternative -- accepting empty publications and raising
the fixed wait -- was rejected: without the completion signal, an
early clear-on-open publication would be taken as the answer, and
every clean file would stall for the full wait.
- Supersedes the quarantine approach of draft PR #1674 (broadening the
xfail to non-Linux); whichever lands second resolves the trivial
conflict on the xfail block.
EclipseJDTLS.is_ignored_dirname hard-ignored directories named target/build/bin/out/classes/dist/lib. Several of these (dist, lib, classes, out) are valid Java package identifiers, so any source file whose package path contained one was dropped from directory traversal and became invisible to the symbol tools (find_symbol, get_symbols_overview, find_referencing_symbols) even when git check-ignore reported nothing.
Per maintainer guidance in #1645, remove the override entirely: the base is_ignored_dirname already skips _ALWAYS_IGNORED_DIRS (VCS/venv/cache/IDE internals), and real build output is already excluded via .gitignore, so the hardcoded list was both redundant and harmful.
Adds a regression test (no Java/JDTLS process required) that fails on main and asserts the formerly-hardcoded names are no longer ignored while _ALWAYS_IGNORED_DIRS still are.
Dominik Jain