Repository navigation
0.7.0 — Stable navigation, explainable search, and project evaluation
0.7.0: Stable navigation, explainable search, and project evaluation
This release makes navigation more consistent across pages, Git revisions, model changes, and multiple local processes. It also adds inspectable retrieval results, project-owned quality evaluations, and automated release attachments.
Fixed
- Search ranking no longer changes with page size or offset. Continuation cursors are bound to the query, filters, synonym configuration, embedding model, and completed index revision.
- Explicit Git ranges analyze both requested snapshots instead of mixing in symbols from HEAD or a dirty working tree.
- ONNX identities fingerprint model files, tokenizer, external weights, and inference settings. Changes to loaded artifacts are rejected so that incompatible vectors are not silently mixed.
- Index initialization and writes are coordinated across processes. Readers use consistent SQLite snapshots; contention produces bounded, actionable errors, and interrupted batches force recovery.
- Context clipping keeps focus lines and reports omitted or individually clipped lines.
- Project synonyms reload on subsequent queries, including file replacement or removal. Invalid settings warn once per content version and recover after correction.
Added
search --explain/ MCPexplain=true: matched query terms, expansion provenance, retrieval-layer ranks and contributions, plus SQL evidence for routed data-access hits.search --cursor <next_cursor>/ MCPcursor: stable continuation with explicit rejection when the index or settings change.evaluate --queries <file>: per-question answers, Recall@k, hit rate, MRR, p50/p95 latency, estimated output tokens, baseline comparisons, and enforceable quality/performance budgets.- A curated set of 15 navigation questions about this repository, with CI-generated evaluation reports.
- A Release pipeline that tests and builds distributions once, validates versions, generates
SHA256SUMS, and reuses the same artifacts for PyPI and GitHub attachments. Upload retries skip identical files and refuse to overwrite conflicting assets. - Expanded cross-platform CI and updated English, Traditional Chinese, and Simplified Chinese documentation.
Compatibility and usage
- An invalid or stale search cursor requires a new search; legacy offsets remain supported without revision checks.
- CLI exit status 4 indicates index contention.
INDEXER_LOCK_TIMEOUTandINDEXER_BUSY_TIMEOUTconfigure bounded waits. - New ONNX identities trigger a full re-index on the first indexing run after upgrading. Use a dedicated model directory and restart the server after changing model artifacts.
- In historical comparisons,
snapshot=afterrefers to the target commit. Therevisionsobject identifies the compared commits. - Evaluation tokens are estimated from rendered character counts, not a model-specific tokenizer. Empty datasets fail, and breached budgets return exit status 3.
Validation and known limitations
- The local Linux/Python 3.12 suite and source-distribution suite pass 105 tests, including ONNX, multi-process, historical-range, pagination, and release-integrity regressions.
- Installed-wheel smoke checks, distribution metadata checks, and actionlint validation passed.
- GitHub CI passed on Linux/Python 3.10, 3.12, 3.13 and Windows/Python 3.12, including distribution installation and the navigation report job.
- The 15 curated repository-navigation questions achieved Recall@5 0.533, below the experimental 0.70 target. The failure report was preserved. These are navigation examples, not collected user questions or an independent semantic benchmark; the CI report job does not enforce a recall threshold.
- Existing static-analysis and dynamic-SQL limitations remain. Historical analysis does not reconstruct external
extra_roots.
Upgrade
pip install --upgrade local-codebase-indexer==0.7.0