release notes · v0.5.10
skylakegrep 0.5.10 — fast scoped file discovery and agent instruction depth
0.5.10 closes the gap between a fast local answer and a useful agent answer. It keeps the 0.5.9 facet-based router, then hardens the paths that real users hit immediately after the release:
- scoped file-location questions should return from filesystem evidence in under a second instead of falling through to semantic cascade;
- foreground search should not block behind hundreds of changed files;
- the footer should report the time the user actually waited;
- LLM agents should know when to ask for a path, snippets, full content, synthesized answers, or JSON.
All examples below use fictional placeholders.
What changed
- Added a generic scoped descriptor + metadata file-discovery lane. When a query names a concrete scope, a target description, and a metadata modifier such as created / modified / opened, skygrep now ranks files inside that bounded root by descriptor coverage, basename evidence, metadata evidence, hidden/generated-path penalties, extension priority, and depth. This is not tied to any private query wording.
- Descriptor-file evidence is final only when the requested depth is a
path/location answer. If the user asks what a file says, requests
--answer, or uses agentic search, semantic retrieval remains alive. - Large incremental refresh work now defers to background indexing instead
of blocking the foreground search. The default foreground cap is 25
pending file changes and can be tuned with
SKYGREP_AUTO_REFRESH_MAX_FOREGROUND_FILES. - Human footer timing now uses command wall time from the
searchcallback instead of stage-local cascade timing. Refresh, routing, scoped discovery, and other foreground work are included, so the footer matches the user's wait much more closely. - Fixed a no-cascade footer edge case where a semantic pool without the expected third counter could raise an index error.
- Documented the public information-depth ladder:
bare
skygrepfor location,--content --detail standardfor snippets,--content --detail full --include ...for scoped deep reading,--answer --contentfor local synthesis, and--json --content --detail standardfor LLM/agent calls. skygrep setupnow writes those depth rules into Claude Code, Codex, OpenCode, Gemini CLI, and Cursor instructions.- Existing
skygrep setupsnippets now refresh automatically after upgrade. Normalskygrep "<query>"andskygrep doctorupdate only the previously managed BEGIN/END block; they never create a new agent integration or modify user-written text outside the marker.
Before vs after
| Query shape | Previous behavior observed during validation | 0.5.10 behavior |
|---|---|---|
| Scoped file-location query with metadata modifier | Could fall through to semantic cascade and wait tens of seconds | path=scoped-file-discovery, semantic skipped, sub-second shell wall time |
| Warm query with many changed files | Could block behind foreground refresh before answering | large refresh deferred to background, search continues |
| Footer timing after foreground refresh | Could show only cascade-stage time | command wall time reported from search entry |
| Existing agent setup snippet after upgrade | Old instruction stayed in place until manual reinstall | managed block auto-refreshes on normal search / doctor |
| Agent choosing output depth | Agent had only broad "prefer skygrep" guidance | setup snippet explains location, content, full-detail, answer, JSON, and explain modes |
Information-depth examples
skygrep "where is the project brief I edited recently?"
skygrep --content --detail standard "what does the API migration plan say about rollback?"
skygrep --content --detail full --include "docs/migration-plan.md" "show the deployment steps"
skygrep --answer --content "summarize the payment retry policy"
skygrep --json --content --detail standard "where is token refresh implemented?"
skygrep --explain "where is token refresh implemented?"
Verification
- Full test suite:
303 passed, 2 warnings, 20 subtests passed. - Targeted release regression slice:
27 passed. - Targeted integration setup suite validates depth guidance, explicit setup refresh, and automatic managed-block refresh.
- Targeted scoped-discovery suite validates descriptor + metadata file discovery returns before embedding for path-depth questions.
- Timing regression validates displayed elapsed time includes foreground refresh delay.
- Generic synthetic CLI smoke:
path=scoped-file-discovery,semantic-skipped, footer0.103s, shellreal 0.29s. - Source privacy scan: required before build.
- Wheel/sdist privacy scan: required before upload.
Compatibility
- No CLI flag was removed.
- Required JSON fields are unchanged.
- No index rebuild is required.
- Existing agent instructions are refreshed only inside the
skygrep setupmanaged marker block. Users who never ran setup are not auto-registered. - The new foreground refresh cap affects latency, not recall: deferred files are picked up by background indexing, and exact/metadata/scoped foreground evidence still answers immediately when sufficient.
Privacy note
All examples in this release use fictional placeholders such as
project brief, API migration plan, and payment retry policy. No real
user prompt, private filename, private folder name, local machine path,
screenshot, or document category is included in the release notes, tests,
docs, wheel, or sdist.