v0.5.0

Added

• lore status --full now probes the Ollama inference runtime with a live /api/embed request,
  catching runner-subprocess failures the binary/daemon/manifest checks miss. Failures render
  through a single render_failure switchboard with structured ProbeError variants and feed the
  CLI, MCP lore_status tool, and the post-install auto-probe identically. (#67)
• SessionStart hook surfaces an actionable warning when the runtime probe fails, including a hint
  line so the agent can suggest the right remedy without a full status dump. (#67)

v0.4.1

Fixed

• SessionStart hook output now reaches the model as hookSpecificOutput.additionalContext instead
  of the terminal-only systemMessage chip, so the pinned-conventions index and meta-instruction
  actually seed the conversation on session start. (#65)

Changed

• PostCompact no longer emits a hook payload. Claude Code's hook output validator rejects
  hookSpecificOutput for this event, leaving only the chip-only systemMessage envelope — which
  never reaches the model and renders as long terminal noise on every /compact. The handler now
  does its dedup-file truncation silently; on-demand pattern injection via PreToolUse continues to
  work. The always-on pinned tier is unavailable post-compact pending a future re-prime mechanism.
  See docs/hook-pipeline-reference.md and ROADMAP.md. (#65)

v0.4.0

Changed

• Trace maintenance now only deletes trace data files (.jsonl / .jsonl.gz) after the retention
  horizon; other files in the trace directory are left alone instead of being swept — applies to
  both lore trace prune and the SessionStart-triggered lazy maintenance pass. (#60)

Added

• MCP add_pattern and update_pattern accept an optional language argument that opts the
  pattern into the structural retrieval gate, warn-and-proceed for unknown tokens. (#63)
• lore status inserts a new Languages: line between Sources: and Last commit: reporting a
  per-language source-count breakdown plus an undeclared bucket; the same data is exposed as
  languages_declared / languages_undeclared on the MCP lore_status tool's metadata fence for
  agent or script consumers. (#58)
• Add opt-in per-hook trace logging under $XDG_STATE_HOME/lore/traces/ with lore trace why query
  CLI and lore trace prune maintenance. (#59)
• lore detects 21 new languages: C, C++, C#, Swift, Kotlin, Shell, Objective-C, Scala, Elixir,
  Dart, Lua, Nix, Terraform, Haskell, Clojure, Zig, Perl, Ruby, Java, Groovy, and PHP. (#61)

Fixed

• Hook-driven language detection now reads both command and description from Bash tool calls;
  previously a populated description shadowed the command and silently no-op'd the structural
  retrieval gate for nearly every real Bash invocation. (#62)
• Path-prefixed Bash invocations like ./gradlew, /usr/local/bin/cargo, and ~/.cargo/bin/cargo
  are basename-normalised before keyword matching, so they resolve to their canonical language
  instead of failing the keyword allow-list. (#62)

v0.3.1

Fixed

• update_pattern and append_to_pattern no longer leave stale chunks in the index when the
  knowledge directory is reached through a symlink (typical on macOS, and on any symlinked or
  bind-mounted setup). (#55)

v0.3.0

Added

• An optional language: frontmatter field declares the languages a pattern targets, so declared
  patterns surface on relevant tool calls even when their bodies omit the canonical token. (#50)
• lore ingest reports a coverage tally distinguishing patterns that declare language: from those
  that fall back to FTS coincidence, and aggregates unknown-token warnings per token across the run.
  (#50)

Changed

• Bash command language inference matches on whole tokens, so bundle install no longer falsely
  infers TypeScript via the bun substring. (#50)
• Shared signals across languages (e.g. npm test) now infer the full applicable set
  ({javascript, typescript}) rather than collapsing to a single arbitrary winner. (#50)
• Knowledge database schema bumped to v4 via a forward-compatible ALTER TABLE migration on first
  open; no lore ingest --force required. (#50)
• The XDG state tier ($XDG_STATE_HOME) is now documented and reachable via a new
  default_trace_dir() helper; XDG path resolution moves to the etcetera crate without changing
  default paths or $HOME-fallback semantics. (#52)

v0.2.0

Added

• add_pattern distinguishes real slug collisions from intentional re-use (with a
  (no title heading) variant for orphan-frontmatter targets), NFC-normalises titles before
  slugifying, and rejects whitespace/newline injection at the write boundary. (#42)
• lore ingest, lore serve, and lore_status warn when the knowledge directory's effective scan
  set is empty, distinguishing filesystem-empty, all-.loreignore-excluded, and missing-directory
  cases. (#41)
• Universal-tagged patterns may declare an applies_when predicate to gate re-injection by tool
  class and Bash command prefix (OR within keys, AND across). (#39)
• min_relevance_universal adds a per-tier relevance floor under [search], defaulting to inherit
  min_relevance. (#39)
• Engine/adapter split moves predicate evaluation, smart-prefix matching, and query extraction into
  agent-agnostic src/engine/, with src/hook.rs as the Claude-Code adapter. (#39)
• lore ingest skips non-UTF-8 filenames during the directory walk and surfaces them on
  IngestResult::errors instead of indexing them under U+FFFD-substituted keys. (#46)

Changed

• lore ingest on a fresh git init with zero commits now prints
  No commits yet — HEAD will be recorded after your first commit. instead of the misleading "No
  previous ingest recorded" wording. (#45)
• Knowledge database schema bumped to v3 via a forward-compatible ALTER TABLE migration on first
  open; no lore ingest --force required. (#39)
• Predicated universal patterns are no longer pinned at SessionStart; they re-inject on every
  matching PreToolUse call via the predicate path. (#40)

Notes

• Additive at every user-facing surface; the v3 schema bump is one-way — a v0.1.x binary against a
  v3 database requires lore ingest --force after re-install.

v0.1.0

First stable release. No user-facing changes since 0.1.0-alpha.1; promoting after end-to-end
pipeline validation. See [0.1.0-alpha.1] below for the full feature list.

v0.1.0-alpha.1

Added

First public release. Lore is a local semantic search engine for software patterns and conventions,
exposed as an MCP server for Claude Code and other MCP clients. Single Rust binary with SQLite,
FTS5, and sqlite-vec compiled in; only runtime dependency is Ollama for embeddings.

Core

• MCP server — lore serve exposes five tools over stdio: search_patterns, add_pattern,
  update_pattern, append_to_pattern, lore_status. Designed for Claude Code's MCP transport but
  works with any MCP client.
• Hybrid search — combines FTS5 lexical search and sqlite-vec vector similarity via Reciprocal
  Rank Fusion. Title and tag matches weighted above body text. RRF scores normalised to 0–1 range.
  FTS5 porter stemming for improved recall. Ollama fallback warning and min_relevance threshold
  guard against poor-quality results. Set hybrid = false in lore.toml to skip Ollama at query
  time.
• DB as sole runtime read surface — pattern bodies live in a patterns table (source_file PK,
  title, tags, is_universal, raw_body, content_hash, ingested_at); SessionStart and
  PostCompact render from the DB instead of re-reading source markdown. Sandboxed read-only agents
  no longer need filesystem access to the patterns directory for the pinned-render path. Agents that
  call write tools still need write access because those tools persist markdown to disk as the
  authoring surface. See docs/architecture.md for the codified invariant.

Ingestion

• Delta ingest — lore ingest only re-indexes changed, added, moved, and deleted files since
  the last-ingested commit (via git diff --name-status), eliminating the Ollama round-trip penalty
  for unchanged files.
• Single-file ingest — lore ingest --file <path> indexes one file without requiring a git
  commit, enabling the fast edit-ingest-search loop for pattern authoring. Respects .loreignore;
  --force overrides.
• .loreignore — gitignore-style exclude file in pattern repositories. Filters files during
  full and delta ingest, with reconciliation when the file changes. Supports negation patterns,
  directory globs, and recursive globs via the ignore crate.

Claude Code integration

• Plugin (integrations/claude-code/) — SessionStart priming with pinned universal patterns,
  PreToolUse hook for relevance-gated pattern injection, PostCompact reset, error hook. Includes
  /search and /coverage-check skills.
• Universal patterns — patterns whose tags: frontmatter list contains universal get
  always-on injection at SessionStart (full body in a ## Pinned conventions section) and bypass
  the PreToolUse dedup filter so they re-inject on every relevant tool call. Additive beyond
  top_k; relevance gate intact. Closes the always-on discoverability gap for process-level
  conventions like commit messages, push discipline, and branch naming. See
  docs/pattern-authoring-guide.md for the new "When to use the universal tag" section.
• /coverage-check skill — automates the Vocabulary Coverage Technique from the pattern
  authoring guide by simulating the PreToolUse hook's own query extraction on synthetic tool calls
  (via lore extract-queries), ingesting a draft pattern, searching in parallel, and iterating on
  edit suggestions until the surfaced-query set stabilises.

Operations and tooling

• Release process — pushing a v* tag to GitHub triggers a workflow
  (.github/workflows/release.yml) that cross-compiles lore for four targets via cargo-zigbuild
  from a single Linux runner (x86_64-unknown-linux-gnu, x86_64-unknown-linux-musl,
  aarch64-apple-darwin, x86_64-apple-darwin), packages each binary into a tarball with both
  license files and the README, computes a SHA256SUMS file for integrity verification, and
  publishes a GitHub Release with the matching CHANGELOG section as the body. The publish step is
  gated by a release GitHub Environment that requires owner approval — push permission alone
  cannot ship a release. CI gains a cross-compile smoke job for the same four targets so
  cross-compile breakage surfaces on every PR. New just release-prep VERSION recipe rotates the
  CHANGELOG and bumps Cargo.toml. Maintainer runbook at
  docs/release-process.md.
• LORE_DEBUG=1 verbose logging via env var; --json structured output for search,
  list, and status for script and agent consumers.
• Security hardening — input limits, transcript path validation under $HOME, bounded tail-read
  (32 KB), deduplication file locking via fd-lock with FNV-1a hashing. See SECURITY.md.
• Documentation — pattern authoring guide, search mechanics reference, hook pipeline and plugin
  reference, configuration reference, release-process runbook.