The code graph indexes every function, import, and dependency. When you change something, it knows exactly what else might break — in under 5 seconds.
30x
Fewer tokens
vs Explore agents — measured on 3 live graphs, 35,758 nodes
9x
Faster
<5 seconds vs 45+ seconds
0
Blind spots
Finds systems agents miss
15s
To build
Full graph from scratch
Each Explore agent spawns with ~20K tokens of overhead. Three agents for a single feature = 60K tokens before any code is written.
Pattern-based file search misses code that doesn't match expected naming. In one test, agents missed 8 existing systems that the graph found instantly.
Multiple agents searching in parallel still takes 45+ seconds and produces 9,000 words of context. Graph queries return in milliseconds.
This is not a best-practice suggestion. It is a PostToolUse hook — the impact query runs before the edit lands. Your agent literally cannot bypass it to write code blind.
Composure parses your codebase with tree-sitter AST analysis. Every function, class, type, import, and test relationship is indexed in a SQLite database. Query it instantly — callers, dependencies, blast radius, entity scope — all without spawning a single agent.
Measured in the field: three live graphs on one development machine index 35,758 nodes and 92,624 edges across 5,207 files, with 5,150 persisted audit findings stored in-graph. The largest — a client operations platform — carries 10,345 nodes and 204 domains over 2,134 files; impact analysis runs against that, not against grep.
Caught in the wild
A reasonable local edit silently breaks something three files away — no agent told either of them. Replit's agent deleted a production database; Cursor wiped PocketOS's. In both cases, the agent had no model of what else depended on what it changed. get_impact_radius runs the blast-radius query before the edit lands. The TDAD paper measured the same class of pre-change impact analysis and found regression rates drop from 6.08% to 1.82% on SWE-bench Verified — not because the AI got smarter, but because it could see what it was touching before it moved.
TDAD paper — pre-change impact analysis on SWE-bench Verified (arXiv) ↗$semantic_search_nodes({ query: "auth" })$→ Found 12 nodes: AuthProvider, useAuth, login(), ...$get_impact_radius({ changed_files: ["src/lib/auth.ts"] })$→ 3 files direct, 7 indirect, 2 untested$query_graph({ pattern: "callers_of", target: "handleLogin" })$→ 12 callers across 8 files$Total: 3 queries, ~500 words, <5 seconds$(vs 3 agents, ~9,000 words, 45+ seconds)Run /composure:initialize or /composure:build-graph. Parses all source files with tree-sitter, indexes entities and relationships into a SQLite database.
Use semantic_search_nodes, get_impact_radius, query_graph, and entity_scope to find code instantly. No agent spawning needed.
The PostToolUse hook incrementally updates the graph after every file write. It's always current — no manual rebuild needed.
With multi-project scope, query across 15+ codebases simultaneously. Find shared patterns, detect version drift, build templates from real data.