v3.1.2

v3.1.2 is an install-path patch release. A first-time user who followed the
README could end up with an MCP entry their client never reads, a Swift or Lua
repository that indexed as unsupported language, or an [all] install that
pulled a test toolchain. This release fixes those paths. No MCP request or
response shape changes, and the index does not need a rebuild.

Fixed

• install.sh writes the file each client actually reads. Claude Code now
  gets a project .mcp.json (the old ~/.claude/.mcp.json and
  ./.claude/.mcp.json targets are not read by Claude Code). VS Code gets
  .vscode/mcp.json with the servers key and type: stdio (VS Code ignores
  mcpServers). Cursor gets a project .cursor/mcp.json; the global
  ~/.cursor/mcp.json is never written, because one fixed
  TREE_SITTER_PROJECT_ROOT there points every project at one repository.
  A detected client whose file does not exist yet gets a new file instead of
  being skipped. Codex CLI users get a printed config.toml block.
• --doctor checks the same locations and reports an unset
  TREE_SITTER_PROJECT_ROOT in the shell as WARN instead of FAIL (the MCP
  client config, not the shell, normally sets it). It flags a VS Code file that
  uses mcpServers and accepts ${workspaceFolder}.
• Swift and Lua grammars ship with the core install. They were optional
  extras, so a plain uvx/pip install answered unsupported language for
  .swift and .lua files. The missing-grammar hint for them now asks for a
  package upgrade.
• README and installation guide (EN/JA/ZH/ES): the automated-install text
  describes what the script does; every quick start includes the one-time
  --full-index step and the expected indexed: true; CLI examples work
  without a persistent install (uvx --from tree-sitter-analyzer ...);
  a Codex CLI block was added; stale "in this PR" / "on develop" wording was
  removed.

Changed

• anthropic is no longer a dependency. No module imports it. This removes
  anthropic, distro, docstring-parser, jiter and sniffio from the
  resolved install.
• [all] means every language grammar plus the MCP extras. It no longer
  installs pytest, ruff, mypy, black, isort, pre-commit or memory-profiler, and
  no longer pins psutil<6. Developer environments use [full] or [dev].
• Core grammar packages carry the same upper bounds as the extras
  (for example tree-sitter-python>=0.23.6,<0.25.0), so a plain install
  resolves the grammar series the test suite runs against.
• install.sh: a config file that fails to parse is now a failure (exit 1)
  instead of a silent skip.

Added

• llms.txt: a non-interactive install guide for AI agents.
• .github/ISSUE_TEMPLATE/agent-trial.yml: a structured report form for an
  agent's install and first-use outcome.
• server.json: title, website, runtimeHint: uvx, and the optional
  TREE_SITTER_PROJECT_ROOT variable for the official MCP Registry; README
  carries the mcp-name ownership marker the registry checks on PyPI.

Migration

• If you installed with [swift] or [lua], nothing changes; both extras are
  kept as aliases.
• If you relied on pip install "tree-sitter-analyzer[all]" for a test
  toolchain, use [full] (or uv sync --extra full in a clone).
• If you import anthropic in your own scripts next to TSA, install it
  yourself; TSA no longer pulls it in.
• If an earlier install.sh run wrote ~/.claude/.mcp.json or
  ./.claude/.mcp.json, Claude Code never read those files. Re-run
  install.sh in the project, or run
  claude mcp add tree-sitter-analyzer --scope project ..., then delete the
  stale file. The same applies to a VS Code mcp.json that uses
  mcpServers.
• Scripts that relied on install.sh exiting 0 on an unreadable config file
  must now handle exit 1.

PyPI: https://pypi.org/project/tree-sitter-analyzer/3.1.2/