v0.2.1: buggo hunt via MCP

buggo hunt is now available via MCP

buggo hunt shipped as a CLI-only command in 0.2.0 - an agent connected over MCP had no way to invoke it, and buggo init's CLAUDE.md/AGENTS.md instructions never mentioned it. This release closes that gap:

• buggo-mcp now registers a second tool, buggo_hunt, alongside buggo_investigate - takes just { repository }, no bug report needed, shares the same per-session call cap.
• buggo init now writes instructions telling the agent when to reach for buggo_hunt (no known bug yet) vs buggo_investigate (a known bug to localize).

No changes to hunt's ranking logic or CLI behavior - this is purely about making it reachable by agents, not just humans running the CLI.

v0.2.0: buggo hunt

buggo hunt: blind correctness-risk triage

New command: buggo hunt --repo <path>. Unlike investigate, it needs no bug report at all — given just a repository, it ranks files by how likely they are to hide an undiscovered bug, using directory structure and a git-derived risk signal (historical fix-commit density, not just raw churn).

Validated before shipping: on 100 real BugsJS bugs, the actually-buggy file fell in the riskiest 10% of candidates 63% of the time by this signal alone (vs. ~10% expected by chance). Showing it to Jev instead of a naive directory-name-only guess raised blind file-level Top-5 triage from 10% to 60% on a 30-bug multi-project sample (McNemar p=0.0007).

Known, disclosed limitation: the signal favors large, frequently-touched "core" files over rarely-touched peripheral files that may still hide a bug. Two correction attempts (size-normalized density, minimum-churn threshold) were tried and measured to be no better — left as an open problem, documented in the README rather than papered over.

hunt stops at file-level triage, same as investigate's suspects — a starting point for you or a coding agent to actually investigate, reproduce, and fix. It does not discover, reproduce, or confirm a bug on its own.

Usage

buggo hunt --repo /path/to/your/repo
buggo hunt --repo /path/to/your/repo --format json

See the README's new "buggo hunt (v0, experimental)" section for the full validation numbers and the known bias.

v0.1.3

Wider frontend file support

Buggo's file scanner previously treated only JS/TS/JSX/TSX/MJS/CJS/PY as candidate source files, so frontend-only bugs (Astro, Vue, Svelte components, plain CSS/HTML) had no real candidates to point at.

• Added .astro, .vue, .svelte, .mts, .cts, .mdx as code candidates, with real symbol extraction (imports/exports/functions) via the existing tree-sitter TS/JS grammars - Astro's frontmatter fence and Vue/Svelte <script> blocks are isolated and parsed directly; .mdx reuses the TSX grammar and tolerates surrounding markdown prose.
• Added .css, .scss, .less, .html, .json as lightweight candidates (selectors, ids/script-srcs, top-level keys - no full AST), same treatment as the existing YAML/Dockerfile config support.
• package.json homepage now points to https://bugo.moraxh.dev.

v0.1.1 - First public release

Buggo's first public release: an AI-native bug investigator that scans a repository and ranks the files most likely responsible for a reported bug, using Jev as a cheap decision layer.

Install

npm install -g @moraxh/buggo
buggo config set-key <your-openrouter-key>
buggo investigate "checkout fails when cart is empty" --repo /path/to/your/repo

Also on npm: @moraxh/buggo (the installed command is buggo; the bare package name was already taken).

What it does

• Scans a repo's production source files (JS/TS/JSX/TSX/Python, plus YAML/TOML/Dockerfile config files) and extracts structural signal via tree-sitter - never full file contents.
• Ranks candidate files with Jev, chunked so large repos (1000+ files) get every candidate a real judgment instead of a heuristic pre-filter dropping most of the codebase.
• Extra context that feeds ranking as observed evidence: --diff <ref>, --recent-changes <n>, stack trace path matching, git recency, and --exclude <file> to re-run cheaper after ruling out a suspect.
• Human CLI (live, Ink-rendered detective view), --format json for agents, and an MCP server (buggo_investigate) with a per-session investigation cap.

Benchmarked

80% Top-1 / 91% Top-5 on 100 held-out real-world bugs (BugsJS), never seen while building the pipeline. Full numbers in the README.

What it doesn't do (yet)

buggo hunt (blind bug discovery), reproduction, and patching are not implemented - v0.1 stops at localization.