unch v0.4.4

unch v0.4.4

v0.4.4 is a small MCP and Codex reliability release. It makes the Codex integration safer to use across real workspaces and refreshes the project documentation around the current setup flow.

Highlights

• MCP tools now accept an optional directory argument, so workspace_status, search_code, and index_repository can operate on the active repository instead of the directory where the MCP server process was launched
• The MCP backend now routes tool calls to a repository-scoped backend and keeps workspace_status cheap: checking another workspace does not create .semsearch
• The Codex MCP setup docs were simplified around the recommended flow: install the npm wrapper, run unch codex install, then restart Codex
• The npm package docs now point users at the same Codex setup flow as the main README and Mintlify docs
• CONTRIBUTING.md now includes a fuller project overview, module map, and development workflow for contributors

Upgrade notes

• If you use Codex with the npm wrapper, update and re-run unch codex install, then restart Codex
• The MCP tool names are unchanged: workspace_status, search_code, and index_repository
• Existing .semsearch indexes remain compatible; this release changes MCP workspace routing, not the index format
• Source installs can use go install github.com/uchebnick/unch/cmd/unch@v0.4.4

unch v0.4.3

unch v0.4.3

v0.4.3 is a small Codex and npm polish release. It focuses on making the npm-installed workflow easier to turn into a working agent setup after installing @uchebnick/unch.

Highlights

• Added unch codex install to the npm wrapper flow, so a user can install the package, register the MCP server, and install the local unch Codex skill with one explicit command
• The Codex setup writes the skill to ~/.codex/skills/unch/SKILL.md, configures longer MCP startup/tool timeouts, and removes the old ~/.codex/prompts/unch.md prompt if it exists
• Refreshed README, npm package docs, and Mintlify installation docs around npm as the recommended install path and Codex skill setup
• Aligned the npm MCP smoke test with the current tools-only MCP server by checking tools/list and workspace_status instead of the removed prompt API

Upgrade notes

• If you installed @uchebnick/unch before this release and want Codex integration, run unch codex install once and restart Codex
• The MCP server still exposes the stable tools workspace_status, search_code, and index_repository
• Source installs can use go install github.com/uchebnick/unch/cmd/unch@v0.4.3

unch v0.4.2

v0.4.2

Release published automatically from tag v0.4.2.

unch v0.4.1

unch v0.4.1

v0.4.1 is the publishable v0.4 release. It includes the v0.4 workflow expansion changes and fixes the release-blocking llama.cpp runtime ABI mismatch that broke the v0.4.0 tag smoke test.

Highlights

• Added --provider openrouter for index, search, and MCP, with unch auth openrouter --token ... for storing API keys outside the project tree
• Added Rust Tree-sitter indexing for functions, methods, structs, traits, impl blocks, enums, modules, and attached docs
• Upgraded the local index schema to isolate snapshots by provider, model, and embedding dimension, so local llama.cpp models and remote OpenRouter models can coexist in the same .semsearch state
• Added unch start mcp, a stdio MCP server exposing workspace_status, search_code, and index_repository
• Added MCP initialize instructions that tell agents to check workspace status first, search before broad file reads, and rebuild the index only when needed
• Added npm wrapper installation support for the CLI and MCP entrypoint
• Updated github.com/hybridgroup/yzma to v1.13.0, fixing first-run local indexing with the current downloaded llama.cpp runtime archives

Upgrade notes

• Local index storage moves to schema version 2. Existing v1 indexes are reset automatically, so run unch index --root . once after upgrading
• OpenRouter credentials can be saved globally in ~/.config/unch/tokens.json; use unch auth openrouter --local --token ... only when you intentionally want repository-local credentials in .semsearch/tokens.json
• MCP uses stdio only in this release. Configure clients with command unch and arguments start mcp, or run unch start mcp directly from the repository root
• If you install from source, use go install github.com/uchebnick/unch/cmd/unch@v0.4.1

unch v0.3.12

unch v0.3.12

v0.3.12 is the recommended stable release for the recent indexing hotfix line.

Highlights

• fixes a native embedding runtime crash that could abort unch index during real-world indexing workloads
• caps embedding input to the smallest runtime-reported bound across NCtx, NBatch, and NUbatch
• keeps the earlier embedding-state cleanup fix for decode returned non-zero: 1
• adds a tagged-release smoke gate that runs a real unch index before publishing release assets
• refreshes the Windows ARM MSYS2 keyring setup in CI to avoid transient package signature failures

Upgrade notes

• If you installed v0.3.11, upgrade to v0.3.12.
• No index rebuild should be required when upgrading from v0.3.10 or v0.3.11.
• To install from source:

go install github.com/uchebnick/unch/cmd/unch@v0.3.12

unch v0.3.11

unch v0.3.11

v0.3.11 is a hotfix release that addresses inference errors (decode returned non-zero: 1) preventing code indexing on wide-context repositories.

Highlights

• Fixed a llama.cpp crash (KV cache overflow) by reliably clearing embedding evaluation state memory between files
• Corrected invalid CGO runtime state that could leak while processing minified JavaScript files sequentially

Upgrade notes

• No index rebuild is required when upgrading from v0.3.10
• If you install from source, use go install github.com/uchebnick/unch/cmd/unch@v0.3.11

unch v0.3.10

unch v0.3.10

v0.3.10 is a small release that changes the project license to Apache 2.0 and fixes crashes during code indexing.

Highlights

• changed the project license to Apache 2.0
• restricted tokenization size to the configured NCtx to avoid context overflow errors
• added filtering for null bytes (\x00) and invalid UTF-8 sequences before the CGO boundary to fix signal arrived during cgo execution crashes
• bumped golangci-lint-action from 8 to 9
• updated the demo GIF in the documentation

Upgrade notes

• no index rebuild is required when upgrading from v0.3.9
• if you install from source, use go install github.com/uchebnick/unch/cmd/unch@v0.3.10

unch v0.3.9

unch v0.3.9

v0.3.9 is an install-reliability and CLI-polish release. It keeps the local-first indexing and search flow intact while making published installers safer, fixing the latest install path so it stays on release assets, adding a simple CLI version check, and simplifying the public embedding-model tuning surface.

Highlights

• shell and PowerShell installers now verify downloaded release assets against the published checksums.txt file before unpacking
• unversioned installer paths now prefer GitHub releases/latest assets instead of unexpectedly falling back to go install
• install smoke coverage now exercises the installer flow on macOS, Windows, Ubuntu, Debian, Arch, and NixOS-like Linux environments
• unch --version, unch -version, and unch version now report the CLI build version, and release binaries inject the release tag into that output
• index and search no longer expose a separate --batch-size knob; --ctx-size remains the public override for model context sizing
• model install metadata now lives in a dedicated internal catalog, reducing duplication between runtime model profiles and install-time model resolution
• install docs now show the current macOS/Linux sudo sh -s -- -b ... flow, the Windows PATH note, and unch --version as a lightweight post-install verification step
• benchmark docs and CI summary flow were tightened up so the checked-in suites are described as unch regression tracking rather than cross-tool comparisons

Upgrade notes

• no index rebuild is required when upgrading from v0.3.8
• no remote index workflow migration is required for repositories already using unch-index.yml
• if you automated unch index or unch search with --batch-size, drop that flag and keep --ctx-size only when you need to override the model default
• you can verify the installed CLI with unch --version and unch --help
• if you install from source, use go install github.com/uchebnick/unch/cmd/unch@v0.3.9

unch v0.3.8

unch v0.3.8

v0.3.8 is a release-confidence patch release. It keeps the v0.3.7 CLI, runtime, and indexing behavior intact while adding regression coverage around first-run downloads, managed yzma installs, and cached model repair flows, and it republishes the current install path under a fresh semver tag for Go module consumers.

Highlights

• adds direct runtime fetch coverage for the GitHub release lookup, temp-file downloader, tar/zip extraction paths, and the yzma archive 404 fallback
• covers automatic known-model downloads and cached GGUF repair flows in the model cache
• exercises managed yzma install resolution and stdout download progress reporting through local test servers and archive fixtures
• raises the portable-package CI coverage after the runtime downloader and archive hardening work that landed in v0.3.7

Upgrade notes

• no index rebuild is required when upgrading from v0.3.7
• if you install from source or use go install, use github.com/uchebnick/unch/cmd/unch@v0.3.8
• if you previously tried the public Go module path at v0.3.7, switch to v0.3.8

unch v0.3.6

unch v0.3.6

v0.3.6 is a benchmark reporting and release automation patch release. It keeps the v0.3.x search and indexing behavior intact while fixing Telegram delivery, tightening cross-platform benchmark comparisons, and polishing the project entry points.

Highlights

• fixes Telegram release notifications for the configured channel target and surfaces Telegram API errors directly in the release workflow
• aligns the release benchmark matrix across Linux, macOS, and Windows by adding macOS, using the same ci suite on Windows, and collapsing per-platform benchmark details in the summary
• reports Windows benchmark CPU info more accurately and adds Telegram News and Telegram Chat badges to the README header

Upgrade notes

• no index rebuild is required when upgrading from v0.3.5
• existing installs can be upgraded through the normal installer path
• keep TELEGRAM_CHAT_ID set to the full Telegram channel id if you use release notifications