Graph-powered PR review with blast radius detection, structural analysis, and implementation planning at 2.75x the efficiency of exploration-based approaches.
2.75x
Fewer tokens
vs Plan-Mode exploration on the same change set
2.6x
Faster
3m 24s vs 8m 50s
39/40
Quality score
scored on the blueprint-quality rubric (v5 progressive template)
0
Agent spawns
Graph replaces exploration
Traditional code review catches typos and style issues but misses structural problems — broken dependency chains, untested callers, permission boundary violations.
Plan Mode spawns 3+ agents for discovery, consuming 204K tokens before writing a single line. That's 99.4% of your budget spent on reading, not writing.
Without forced structure, AI plans skip preservation boundaries, miss edge cases, and produce "Approach" paragraphs instead of per-file implementation specs.
Structural findings gate the commit. This is not a code style suggestion — project-authored rules are enforced at the hook layer. The model cannot mark a finding resolved without fixing it.
Composure's review maps every change against the code graph. Every review checks: blast radius (callers reached by the change) · untested callers · decomposition thresholds · a diff-scoped security scan (Semgrep + secrets sweep + graph-weighted exposure ranking) · your own project-authored rules. Write a rule once — every review enforces it automatically. Blueprint planning uses the same graph for instant discovery, then progressive refinement with a structured template forces detailed per-file specs. The result: plans that catch design flaws during planning, not implementation.
See it in the field. In a recent client campaign, the review gate surfaced BLOCKING findings mid-run — two security findings caught by dedicated reviewer agents and fixed in-session, with the loopback taking four minutes of active time. The session record is public:
$/composure:review$→ Changes: src/lib/auth.ts, src/hooks/useAuth.ts$→ Impact: 3 files direct, 7 files indirect$→ Risk: Medium (auth module, 12 callers)$→ Untested: handleLogin() — 12 callers, 0 tests$→ Decomposition: useAuth.ts at 180 lines (threshold: 200)$Blueprint quality evolution:$ v1 (old template): 28/40$ v3 (new template): 34/40$ v5 (progressive): 39/40 — surpassed Plan ModeBlueprint identifies whether this is a new feature, enhancement, refactor, bug fix, or migration — each type gets different questions and analysis depth.
Instead of spawning agents, query the code graph for related files, callers, and dependencies. 11 queries vs 76 agent calls — 6.9x more efficient.
Checkpoints at each step let you shape the plan through conversation. By the time the blueprint is written, most questions are already resolved.
The template forces per-file specs with exact conditions, signatures, and preservation boundaries. Writing this level of detail catches design flaws that vague plans miss.