v0.2.2

Changelog

• 2ca87aa chore(release): v0.2.2 — self-installing plugin + /bm-statusline
• 7664d28 ci(goreleaser): keep-existing release mode + refresh brew formula desc
• 2372fa3 docs(readme): update statusline example to the redesigned format

---

Released by GoReleaser.

v0.2.1

Changelog

• 38f340b chore(plugin): strengthen marketplace entry metadata
• 2da0994 chore(release): v0.2.1 — fix Claude slash commands, harden publish
• 60f3508 ci: install protoc in nightly hardening job
• d126597 fix(plugin): place Claude skills/commands at plugin root + add PRIVACY.md

---

Released by GoReleaser.

v0.2.0

Changelog

• e602dde chore(deps): drop kreuzberg git-pin, swap to crates.io =5.0.0-rc.18
• 529b76f chore(meta): align keywords + description across all surfaces
• b96ac99 chore(release): v0.2.0 — crates.io deps, statusline redesign, README rewrite
• 0576533 docs(readme): full positioning rewrite as AI context layer
• d0ed161 feat(plugin): redesign statusline + bundle agent-facing skills

---

Released by GoReleaser.

v0.1.1

Changelog

• 1e13b44 chore(release): v0.1.1 — npx routing fix + re-enable homebrew
• c11e6bc ci: trim per-PR matrix to default features + skip homebrew tap
• c888322 fix(npm): ship Node shim so npx routes through wrapper

---

Released by GoReleaser.

v0.1.0

0.1.0 — 2026-06-15

Initial public release. First minor wipes the schema (RELEASE_MINOR=1)
so any pre-tag .basemind/ cache rebuilds on next scan — intentional.

Feature-complete code-map server with the kreuzberg document tier surface
(reranker, keywords, NER, summarization, language detection, TOON output) and
schema-driven config across TOML / CLI / MCP / env vars, distributed across
all major coding-agent harnesses (Claude Code, Codex, Cursor, Gemini, Factory
Droid, OpenCode, Copilot CLI, and the generic MCP path) plus npm / PyPI /
crates.io.

Added

• basemind scan — rayon-parallel scanner that indexes a workspace into
  content-addressed msgpack blobs (.basemind/blobs/) plus a Fjall-backed
  inverted index (.basemind/views/<view>/index.fjall/). Two extraction tiers
  ship: L1 outlines (symbols, signatures, imports, docs) and L2 call sites;
  L3 structural hash available for symbol-history diffing.
• basemind serve — stdio MCP server (rmcp) exposing the full code-map
  and git-history tool surface (outline, search_symbols, find_references,
  find_callers, list_files, dependents, repo_info, status,
  symbol_history, working_tree_status, recent_changes,
  commits_touching, find_commits_by_path, diff_file, diff_outline,
  hot_files, blame_file, blame_symbol).
• Dynamic 300+ language coverage via
  tree-sitter-language-pack.
  Hand-written .scm overrides ship for Rust, Python, TypeScript, TSX,
  JavaScript, Go. Other languages for which TSLP ships a vendored tags.scm
  (kotlin, csharp, swift, cpp, scala, solidity, lua, …) get best-effort
  symbol + call extraction via the fallback adapter that rewrites
  GitHub-standard @definition.* / @reference.call captures into
  basemind's @symbol.* / @call.* shape.
• Schema-driven config across TOML / CLI / MCP / env vars. Rust types in
  src/config/ derive schemars::JsonSchema; the snapshot at
  schema/basemind-config-v1.schema.json is regenerated from those types and
  asserted by tests/config_schema.rs. Adding a config field lights up all
  four surfaces via #[command(flatten)] (clap) and #[serde(flatten)] (MCP
  params). Precedence: MCP > CLI > env > TOML > defaults, with per-field
  provenance tracking via src/config/source.rs.
• Document tier — search_documents MCP tool over Lance-backed embedding
  index with kreuzberg ingestion. Per-query overrides on every documents.*
  and llm.* setting, plus entity_category and keywords_contains
  post-filters.
• TOON wire format for MCP responses via
  documents.output.format = "toon" (or --documents-output-format toon).
  Round-trip parity with JSON asserted in tests/mcp_smoke.rs.
• Language-aware ingestion. documents.language.{auto_detect, min_confidence, detect_multiple} flows into kreuzberg's
  LanguageDetectionConfig and the chunking tokenizer. ISO 639-3 codes (e.g.
  "fra") surface on FileMapDoc.detected_languages.
• Cross-encoder reranker as a post-step on search_documents. Off by
  default; documents.reranker.{enabled,preset,top_k} opts in. Preset is
  validated upfront against kreuzberg::get_reranker_preset; reranked index
  is bounds-checked before reorder.
• Keyword extraction (YAKE / RAKE) and named entity recognition at
  extract time. New tail fields on FileMapDoc (keywords, entities)
  with #[serde(default)] — blob-compatible with prior blob shapes
  (asserted by tests/schema_bump.rs).
• Extractive + abstractive summarization via documents.summarization.
  Abstractive routes through liter-llm with the new top-level [llm]
  section (model in provider/model form, api_key, base_url, temperature,
  timeout, retries, max_tokens). NER backend llm now wires the resolved
  LlmConfig.
• SecretString newtype + ApiKey enum (Literal | Env | Unset).
  Secrets mask to "<redacted>" in Debug / Display and across
  Serialize (including the toml→serde_json validation round-trip).
• Real-OSS hardening harness (tests/harden.rs, ./scripts/harden.sh)
  — clones 8 upstream repos (ripgrep, tokio, typescript, react, django,
  requests, gin, ripgrep-shallow), exercises every MCP tool against each,
  and pins canary lower bounds (tokio: find_references("spawn") >= 200,
  django: find_references("get") >= 200,
  react: search_symbols("useState") >= 20,
  ripgrep-shallow truncation surfaces).
• Schema sync to release version — RELEASE_MINOR in src/version.rs
  drives both INDEX_SCHEMA_VER and the blob SCHEMA_VER. Minor-release
  bumps wipe .basemind/ on next scan; patch releases stay compatible.
• Distribution: brew install Goldziher/tap/basemind,
  npm install -g basemind, pip install basemind,
  cargo install basemind --locked. Precompiled binaries on GitHub
  Releases for {x86_64,aarch64}-{linux-gnu,apple-darwin} and
  x86_64-pc-windows-gnu.
• Per-harness install matrix — manifests + skill bundle for every major
  coding-agent harness, all bumped in lock-step by
  task release:sync-version VERSION=…:
  • Claude Code — .claude-plugin/plugin.json + marketplace.json +
    statusline.sh (live one-line summary of the indexed map with true-color
    brand mark, freshness dot, and call/token-savings telemetry from
    .basemind/telemetry.jsonl). Install via
    /plugin marketplace add Goldziher/basemind then
    /plugin install basemind@basemind.
  • Codex CLI / App — .codex-plugin/plugin.json with full interface
    block (displayName, category Developer Tools, capabilities,
    defaultPrompt, brandColor). scripts/sync-to-codex-plugin.sh mirrors
    into a fork of the openai/plugins marketplace.
  • Gemini CLI — root-level gemini-extension.json with contextFileName
    and mcpServers. Install via
    gemini extensions install https://github.com/Goldziher/basemind.
  • Cursor — .cursor-plugin/plugin.json.
  • OpenCode — published as the
    basemind-opencode
    npm package. Install via
    { "plugin": ["basemind-opencode@latest"] } in opencode.json. Skills
    are bundled into the tarball; the plugin shim does dual-mode resolution
    (bundled-path-first, repo-root-fallback) so monorepo dev and npm install
    both work without duplication.
  • Factory Droid / GitHub Copilot CLI — reuse the existing
    .claude-plugin/marketplace.json per their published patterns.
  • Every manifest carries the same canonical user-facing description and
    crates.io-capped 5-keyword set (mcp, tree-sitter, code-map,
    scanner, indexer).
• Release pipeline — .github/workflows/publish.yaml publishes on
  every v* tag: GitHub release assets (cross-compiled binaries via
  goreleaser + zig), crates.io, npm × 2 (basemind binary wrapper +
  basemind-opencode OpenCode plugin), and PyPI. Per-registry idempotent
  via existing-version detection; OIDC-driven trusted publishers on all
  four registries.

Changed

• tree-sitter-language-pack: =1.9.0-rc.27 → =1.9.0-rc.45. Hierarchical
  data_extraction + 17 data formats; rc.40–45 are CI/codegen-only.
• kreuzberg bumped to the reranker / LLM API surface
  (=5.0.0-rc.7 baseline → published rc covering reranker + LLM at publish
  time).
• alloc-stdlib = "=0.2.2" pin lifted (no longer binding after lock
  re-resolve; single alloc-no-stdlib 2.0.4 in the tree).
• src/mcp/types.rs + src/mcp/helpers.rs split into _documents.rs
  siblings to stay under the 1000-line cap.

Performance

• Harden 8/8 green across ripgrep / tokio / typescript / react / django /
  requests / gin / ripgrep-shallow. All canaries pass. Per-repo scan times
  within baseline: typescript 21.7 s (81 324 files), tokio 0.2 s (859
  files), django 2.5 s (7 061 files), react 2.2 s, requests 0.7 s, gin
  1.0 s, ripgrep 4.0 s, ripgrep-shallow 0.16 s. All 25 MCP tools clean
  across all repos.
• search_documents post-processing releases the store read-lock before
  blob I/O; ahash::AHashMap / AHashSet on the post-filter path.

v0.0.1

• initial