release notes · v0.6.0

skylakegrep 0.6.0 — faster indexing and authoritative agent evidence

0.6.0 is a performance-and-accuracy release for agent workflows. It removes per-file embedding request overhead, keeps repeated daemon calls responsive, and makes the evidence behind agent confidence and synthesized answers more explicit.

The speed work does not relax retrieval quality. Closed-loop fixtures now enforce path-only locate semantics, retrieval-only queries cannot claim lexical confidence they did not earn, and local answer synthesis keeps current policy/reference documents separate from unrelated plans or version snapshots.

What changed

• Full, incremental, explicit-file, and watch indexing batch chunks across file boundaries before requesting embeddings. The default batch is 64 and is clamped to 1..512 through SKYGREP_INDEX_BATCH_SIZE.
• Watch mode refreshes file embeddings and removes stale chunks when a tracked file becomes empty.
• skygrep doctor checks optional dependencies without importing heavyweight reranker modules.
• skygrep serve binds before optional reranker initialization. Reranker warming is explicit and asynchronous through --warm-reranker.
• Agent summaries expose confidence_basis as either lexical-and-retrieval or retrieval-only; non-lexical queries receive no synthetic lexical coverage credit.
• Candidate-recall and source-prior handling are shared across the agent path, including a decisive but reversible prior against unnamed historical version snapshots.
• --answer filters unrelated snapshots, roadmaps, and planning notes when a current living policy/reference document leads. Exact-version queries and explicitly named supporting documents remain available.
• The closed-loop benchmark contract now represents locate tasks as path-only tasks and preserves numeric fixture evidence.
• GitHub Actions use Node 24-compatible actions/checkout@v5 and actions/setup-python@v6.

Measured performance

All comparisons below were run on the same repository and codebase family. Indexing, hot CLI operations, and closed-loop agent quality are reported as separate receipts.

The adaptive comparison also used 9.82x less context and had 3.34x lower estimated agent elapsed than the raw-rg policy. Raw rg remains the correct tool for exact regex and exhaustive raw lexical output.

Accuracy and evidence boundaries

The release treats speed and correctness as independent gates:

• deterministic file/chunk ordering is preserved across embedding batches;
• exact-version queries retain their matching historical document while other versions of the same document are excluded from current-policy synthesis;
• historical tasks remain unchanged when no living authority leads;
• confidence_basis states which evidence family supports the summary;
• a live release-policy answer cited only docs/RELEASING.md:106-121 after unrelated version and planning documents were filtered;
• original ranked sources remain visible below synthesized answers.

Compatibility

• Existing indexes are byte-compatible; no rebuild is required.
• The top-level agent JSON value remains a list of result objects. New fields are optional.
• Daemon startup no longer warms the optional reranker by default. Workloads that rely on later reranked queries can opt in with skygrep serve --warm-reranker.
• Exact regex/raw-output workflows should continue to call rg directly.

Verification

• Full implementation suite: 383 passed, 20 subtests passed with two existing Click deprecation warnings; the isolated Python 3.9 release environment also completed 383 passed with one environment-only LibreSSL warning.
• Targeted answer/document-policy suite: 7 passed; the Python 3.9 enrichment lifecycle regression suite completed 4 passed.
• Pull-request CI: Python 3.10, 3.11, and 3.12 all passed.
• Post-merge master CI: Python 3.10, 3.11, and 3.12 all passed.
• Medium closed-loop benchmark: 8/8 complete, 100% path coverage, 100% evidence coverage, 100% sufficiency, no failures.
• Live --answer release query completed with the canonical release document as its only source.

Privacy

Source, documentation, and the actual GitHub PR diff passed the privacy gate. Public examples use fictional paths and generic prompts. No real local prompt, private filename or folder, absolute local path, screenshot, document category, name, email, credential, token, or machine-specific identifier is included. Wheel and sdist artifacts are scanned again after the release build.

Known follow-ups

• PyPI Trusted Publishing still needs the one-time project-side publisher registration for owner danielchen26, repository skylakegrep, workflow release.yml, and environment pypi. Until that account setting is added, releases use the existing authenticated twine path after the same build, metadata, artifact, and privacy gates. This affects automation only, not the package contents or availability.