Added v3 graph snapshots with full BLAKE3 payload identities, hydrated search and dependency indexes, and transparent v0.9-v2 loading.
Added a reproducible performance gate requiring 20% faster 500-file indexing and warm exact search than v0.9.0.
Added release-asset checksum verification to the Unix and Windows installers.
Added cross-platform, website, dependency-audit, and installer checks to CI.
Fixed
Detect same-size source changes even when file modification times are preserved.
Preserve line endings and executable permissions during atomic line edits.
Bound newline-delimited MCP messages, headers, and agent-facing result counts.
Report one deterministic finding per dependency cycle and downgrade cycles contained within one Rust module family.
Changed
Centralized shared CLI/MCP read, patch, create, and audit workflows plus MCP graph lifecycle persistence.
Reduced index memory and CPU overhead by removing duplicate content caching and deduplicating word and trigram data.
Parallelized parsing and search-index preparation for projects with at least 64 files.
Limited exact-word hit materialization to the requested result count.
Pinned every external GitHub Action to an immutable full-length commit SHA.
v0.9.0 - 2026-06-30
Added
Added agent retrieval and CLI output format benchmark coverage.
Added the Astro and Tailwind website with Guide, MCP, Skill, Benchmark, and Changelog pages.
Changed
Improved agent-oriented TOON output and audit ranking.
Improved brief source ranking for more useful task context.
Reduced structured text token overhead with flattened success metadata, shared path roots, direct path arrays, trimmed search text, conditional next steps, and triage-first audit output.
Refreshed the public benchmark docs and site page to cover retrieval, CLI output format, edit safety, MCP session, and maintenance suites.
v0.8.0 - 2026-06-25
Added
Added a canonical MCP TOOL_SPECS table plus generated skill/SKILL.md and docs/tools.md workflows.
Added xtask gen-skill for regenerating MCP tool documentation from TOOL_SPECS.
Changed
mcp::tools() now consumes TOOL_SPECS, and MCP tools/list descriptions use short summaries.
Tests
Added invariant tests for MCP tool specs and generated tool-list shape.
v0.7.0 - 2026-06-22
Changed
Rewrote all 22 MCP tool descriptions in src/mcp.rs::tools() from terse one-liners into 2-3 sentence guidance covering when to use each tool, what it returns, and key arguments, helping agents pick the right tool more reliably.
Synced the same descriptions verbatim into the MCP tools table in skill/SKILL.md (also added the previously missing symbol_search, reindex, and clear_index rows and reordered rows to match src/mcp.rs).
v0.6.7 - 2026-06-21
Fixed
Resolved default graph paths relative to the project root so lexa index /path/to/project writes the graph under that project by default.
Prevented --no-graph edit, create, index, and watch flows from persisting snapshots.
Made persisted patch and create fail before mutating files when no graph exists, avoiding accidental one-file snapshots.
Made audit fail on empty indexes instead of reporting a clean result.
Made watch fail startup when an existing graph cannot be loaded.
v0.6.6 - 2026-06-18
Fixed
Ignored hidden project paths during MCP watcher refreshes so .lexa/mcp.log and other internal files do not trigger repeated graph updates.
v0.6.5 - 2026-06-17
Added
Added lexa mcp --log-file <path> to append MCP startup and runtime diagnostics to a plain-text log file.
v0.6.4 - 2026-06-12
Fixed
Resolved TypeScript ESM runtime import specifiers such as ./module.js to matching .ts, .tsx, .mts, and .cts source files.
Resolved local Vite-style resource imports with query or hash suffixes such as styles.css?url.
v0.6.3 - 2026-06-11
Fixed
Rejected oversized MCP stdio messages before allocating request bodies.
Made snapshot loading replace existing engine state instead of merging stale indexes.
Changed
Removed confirmed unused engine and walker helpers.
v0.6.2 - 2026-06-10
Tests
Expanded regression coverage for MCP read, patch, create, status, recent, and changes flows.
Added focused coverage for edit/create safety paths, pipeline output behavior, output helpers, CLI line ranges, and parser fixtures.
v0.6.1 - 2026-06-04
Fixed
Fixed a clap argument-id collision that caused lexa upgrade to panic when the global --version flag and upgrade positional version argument shared the same internal id.
v0.6.0 - 2026-06-04
Added
Added top-level symbol_search for fuzzy symbol discovery when exact symbol_defs is too strict.
Added reindex and clear_index MCP tools for explicit graph recovery.
Added architecture cycle detection to structural audit output.
Added files filtering by path, glob, language, and line-count bounds.
Changed
Removed the legacy MCP pipeline.query argument; pipeline now accepts pipeline or steps.
Made brief explicit about its scope as a context bundle for symbols, paths, and scoped keywords rather than natural-language QA.
Improved brief ranking and body extraction for relevant definitions.
Made lexa index print a lightweight branded banner in interactive terminals.
Moved CLI upgrade/version-check code into a dedicated module and centralized shared output formatting.
Fixed
Prevented invalid graph snapshots from silently loading as an empty index and producing misleading command results.
Made audit refuse to run when no files are indexed.
Added header-first snapshot validation so incompatible graph versions fail before payload decoding.
Cleaned outline output by keeping imports out of the symbol list and improving missing-file/config error messages.
Improved JSON outline classification consistency and removed an avoidable package manifest parser unwrap.
Named brief scoring weights to make future ranking changes easier to review.
Tests
Added regression coverage for snapshot header validation, graph-loading behavior, pipeline schema cleanup, fuzzy symbol search, outline import filtering, and parser edge cases.
v0.5.1 - 2026-06-04
Fixed
Resolved local asset imports for SVG/PNG provider logos and other known asset files without indexing binary bytes.
Returned clear metadata stubs from read for known binary assets instead of reporting them as missing.
Normalized TypeScript imports in outlines and dependency data to module specifiers such as ./assets/logo.svg.
Avoided TypeScript outline false positives where exported object or string values could be misclassified as imports.
Made Unix upgrades install through a staged binary and atomic move, avoiding macOS Killed: 9 failures after in-place replacement.
Improved
Made brief prefer relevant symbol definitions before generic snippets or call sites.
Improved brief natural-query handling with identifier, path, and phrase candidates.
Ranked callable definitions such as useTerminalSession and createProjectAgent above related type aliases when both match the same concept.
Bounded large brief symbol bodies to 120 lines.
Tests
Added regression coverage for asset import resolution, metadata-only asset reads, TypeScript import normalization, brief definition ranking, phrase/path-based brief lookup, and large symbol body truncation.
Verified the Unix installer path with a temporary install directory.
v0.5.0 - 2026-06-04
Changed
Made MCP structured content opt-in to reduce duplicated tool output by default.
v0.4.2 - 2026-06-04
Changed
Clarified audit verification limits and reinforced that structural audit output does not replace build, typecheck, lint, or test verification.
v0.4.1 - 2026-06-04
Fixed
Detected unresolved local TypeScript imports.
Fixed Windows installer ZIP layout handling.
Read MCP stdin framing as bytes to avoid non-UTF-8 input failures.
v0.4.0 - 2026-06-03
Added
Added MCP graph freshness checks and watcher support.
v0.3.0 - 2026-06-03
Added
Added the first audit command implementation.
Added scoped audit strict mode.
Added audit configuration.
Added dead-code audit candidates.
Improved
Refined audit reporting and release workflow behavior.
v0.2.0 - 2026-06-03
Added
Added the Lexa binary upgrade command.
Improved
Improved import dependency resolution.
Restricted release workflow execution on pull requests.
