Repowise indexes a repository once — symbols and the dependency graph, git history, tests, documentation, and architectural decisions — and then answers questions about it for both developers and coding agents, so an agent does not rediscover the same structure on every task. The index is built and served locally, and the core analysis needs no model key.

Agents reach the index through ten task-shaped Model Context Protocol tools. Each tool takes several targets in one call and returns complete context, rather than one file or one symbol per call. Optional hooks push context into a session as it becomes relevant: the architectural decision governing a file the agent is editing, a warning on a file with a run of recent bug fixes, or a briefing at session start. Repowise also generates CLAUDE.md and AGENTS.md from the index. Agent integrations are described in two tiers — a Full tier with skills, slash commands, a managed instructions file, hook-level interception, and transcript mining, and a Good tier with MCP tools and the configuration to reach them.

For changes, it traces symbol-level blast radius, historical co-change partners, ownership, missing companion files, breaking contracts, and the tests a diff actually exercises. For maintenance, it produces a 1–10 health score separated into defect risk, maintainability, and performance, detects dead code, and generates graph-aware refactoring plans. Workspaces extend the same model across several repositories to match backend and frontend contracts and enforce architecture rules through one MCP endpoint.

Published measurements include a 31.6% reduction in agent output tokens over 43 questions on a pinned django/django commit (3.8 versus 7.2 tool calls, p<0.0001), a get_context retrieval payload of 393 tokens versus 13,984 across 30 Flask commits, 2.3× more defects surfaced at the same 20%-of-lines budget across 2,770 files (p=0.003), 0.876 file coverage on ContextBench across 42 sealed Python and Go instances, and 0.74 ROC AUC for defect prediction across 21 repositories and 9 languages. Graph, risk, health, test, dead-code, and pull-request analysis run without model calls; generated prose is optional. 25 languages are parsed to an abstract syntax tree, with 39 placed on a five-rung support ladder.

Features

One local index: code and dependency graph, git history, tests, documentation, and decisions built in a single pass
Ten MCP tools: task-shaped tools that take several targets per call instead of one entity at a time
Session hooks: decisions, bug-fix warnings, and session briefings pushed into an agent session as they become relevant
Generated instruction files: CLAUDE.md and AGENTS.md written from the real index
Change intelligence: symbol-level blast radius, co-change partners, ownership, missing tests, and breaking contracts
Code health: a defect-validated 1–10 score split into defect risk, maintainability, and performance, down to the file
Dead code and refactoring: unused code detection and graph-aware refactoring plans an agent can act on
Pull-request bot: change risk and review analysis posted on a pull request, linking to a fuller page
Workspaces: contract matching, breaking-provider detection, and architecture rules across several repositories
VS Code extension: health in the gutter, callers and ownership on hover, refactoring plans as CodeLens, and MCP registration
Local by default: analysis stays on your infrastructure and no API key is needed for the core layers

Discussing Repowise· 0

0 threads · Show & tell