The Codex sandbox on Windows denies the pipes that the daemon and worker
process use, so searches waited 120 seconds and then failed. Detect that and
run the search in-process.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Load both models in bfloat16 when a CUDA GPU supports it and keep float32 on
CPUs. Document the daemon versioning and the indexed file types.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Drop the trailing space after "Query:" as in the model card and include the
prompt in the query cache key.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Code after a top-level closing brace belonged to the preceding function, for
example shell loops labeled as quote_toml. Attribute it to the module.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add HTML, CSS, SCSS, XAML, SQL, XML, project, Razor, Kotlin, Gradle, and
module JavaScript files with SQL and Kotlin symbols, and rebuild the index.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Derive the daemon address and start lock from a hash of server.py, so an
updated server never reuses a daemon still running older code.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
os.execv does not quote arguments on Windows, so the hand-off to the virtual
environment split every multi-word --query and search always failed. Run the
runtime with subprocess on Windows instead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each server.py search/rebuild invocation was a fresh process that reloaded
the embedding and reranker models from disk before answering a single
query. Add an optional daemon (server.py daemon) that keeps both models
resident in memory and serves repeated queries, across repository roots,
over a local socket/pipe.
search and rebuild now try the daemon first, spawn one on a cache miss
(guarded by an exclusive lock file to avoid duplicate spawns), and fall
back to the original one-shot in-process worker if the daemon cannot be
reached or started, so behavior is unchanged when the daemon is
unavailable. The daemon exits on its own after 30 minutes of inactivity.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Pin the Qwen embedding/reranker models to specific commit revisions (and
sentence-transformers to the exact version already in use) so index
builds are reproducible instead of floating on each model's main branch.
Also move the encoder and reranker instances to module-level singletons
shared by every Index, instead of one pair per instance, so a process
serving multiple repository roots only loads each model once.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Repositories with more than 2000 indexable files silently lost the
remainder during indexing. Make the cap configurable via
SEMANTIC_RETRIEVAL_MAX_FILES (default unchanged) and print an explicit
stderr warning naming the repository when it is reached, instead of
truncating without any signal.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Julian lechner
fraujulian