chunkhound/chunkhound v6.1.0

ChunkHound v6.1.0

on GitHub

ChunkHound v6.1.0

This release adds opt-in usage analytics, native embedding providers in the Rust pipeline and the Requesty LLM provider. It also includes Windows CLI launch fixes, wider Linux wheel compatibility and a batch of indexing reliability fixes.

Highlights

• Opt-in usage analytics — per-user usage events with identity anonymization, off by default.
• Native embedding providers — OpenAI-compatible (incl. Azure) and VoyageAI embeddings run inside the Rust pipeline.
• Requesty LLM provider — new OpenAI-compatible provider.
• Windows .cmd/.bat CLIs — npm-only Claude Code, Codex and OpenCode installs now work for research.
• Linux wheels on glibc 2.31+ — chunkhound-native is tagged manylinux_2_31.

Added

• Opt-in usage analytics — ChunkHound can record per-user usage events (CLI commands, MCP tool calls, LLM and embedding/rerank calls with token counts) and upload batches to an S3/MinIO-compatible bucket. Off by default; enable with analytics.enabled or CHUNKHOUND_ANALYTICS__*. analytics.anonymize (full, hashed, anonymous) controls how the user is identified. Free-text content such as queries and URLs is recorded as null unless analytics.save_sensitive_data is set. S3 write credentials come from .chunkhound.json or CHUNKHOUND_AWS_ACCESS_KEY_ID/CHUNKHOUND_AWS_SECRET_ACCESS_KEY, never ambient AWS_* variables.
• Native embedding providers — The Rust pipeline now calls OpenAI-compatible (including Azure) and VoyageAI embedding endpoints itself, with retry, token-aware batching, and the same dimension, TLS-verify and context-length handling as the Python providers. Other providers keep using the Python path.
• Requesty LLM provider — Requesty is supported as an OpenAI-compatible LLM provider (provider: requesty or CHUNKHOUND_LLM_PROVIDER=requesty).

Fixed

• Windows .cmd and .bat CLIs launch through cmd.exe — An npm-only Claude Code, Codex, or OpenCode install can run research. A spaced shim path that cannot be shortened is rejected. Newlines, percent signs, quotes, and a command operator in an argument with no spaces are rejected. Addresses #118. The database WinError 183 in that issue is unchanged.
• OpenRouter structured calls with reasoning models — ChunkHound now detects whether a model accepts the payload that disables reasoning for structured output, retries once without it when rejected, and caches the result per model, so structured calls no longer fail on models that reject it.
• Linux native wheels run on older distributions — chunkhound-native Linux wheels (x86_64 and aarch64) are now built in a pinned Ubuntu 20.04 container and tagged manylinux_2_31, instead of manylinux_2_34. They install on glibc 2.31+ hosts such as Ubuntu 20.04, not only glibc 2.34+.
• Orphaned rows after forced reindex — Rows of deleted files are now cleaned up on a forced full reindex under the Rust pipeline.
• Unchanged re-index no longer rebuilds the HNSW index — A run with nothing to write, delete or compact leaves the vector indexes in place instead of dropping and rebuilding them.
• Responsive server during index maintenance — HNSW maintenance runs off the event loop, so MCP responses and file monitoring are not blocked during rebuilds.
• Faster missing-embedding backfill — generate_missing_embeddings uses targeted queries instead of full-table scans.
• Rerank retries — The single-batch rerank path now retries transient failures like the multi-batch path.
• Known file types no longer fail the NUL-byte check — Files of known types are not skipped as binary because of embedded NUL bytes.
• Path lookup safety — Logical path lookup rejects .. escapes out of the project root.
• MCP regex search with commit inputs — search with type="regex" now warns that commit_range, commit_hash and last_n_commits are ignored, instead of silently returning whole-index results.
• Invalid-argument errors — Research service, CLI search and deep_research validation failures raise ValueError rather than generic errors.
• Embedding provider registration failures are logged — A provider that fails to register is now reported instead of disappearing silently.
• C# and MATLAB parsing — Fixed an import error that broke the C# and MATLAB concept mappings.
• MCP SDK compatibility — The MCP SDK dependency is constrained to the supported 1.x API.
• DuckDB — Sequence names in nextval() calls are parameterized.

Changed

• Claude Code CLI prompt delivery is configurable — One mode cannot fit every machine, so llm.claude_append_system_prompt chooses file (default, every OS) or inline. file uses --append-system-prompt-file: instructions stay in the system prompt, off the Windows command line, and beside a wrapper's --append-system-prompt. inline is for a wrapper that uses the file flag, or Claude Code older than 2.0.34. Those files use the system temp directory. llm.claude_temp_dir overrides it for a special case such as Docker or a limited user. See Claude Code CLI.

Full changelog: https://github.com/chunkhound/chunkhound/blob/main/CHANGELOG.md