Problem
A coding agent about to change a function does not know what a senior maintainer would: which routes reach it, what it writes, and what migrations and past commits say about it.
What I built
codegraph gives the agent a preflight packet before it touches a symbol: route, security checks, idempotency, data writes, migrations, related handlers, invariants, git and pull-request history, and source anchors. Every claim is tagged with where it came from (static analysis, framework, schema, git, repository evidence or heuristic), and a fact it cannot support is left out rather than guessed. It extracts Python, TypeScript and JavaScript with Tree-sitter, stores the graph in SQLite, and exposes its tools over MCP.
code-council lets Claude, Codex and local agents pass proposals and critiques back and forth while planning. A judge mediates convergence instead of taking a simple vote, because a single model's plan tends to miss tradeoffs.
How it was verified
In an early evaluation, five of five packets contained no false claims, each under 2,500 tokens. On a production codebase, a packet surfaced earlier security-hardening intent that justified a narrow two-line edit and ruled out a wider change the evidence did not support.
Outcome
In active development.
Stack
- Tree-sitter
- SQLite
- MCP