AI architectural audit, made tractable by ~10× cache-driven cost reduction — multi-agent debate at single-agent prices.
A 15-cell jury of monomaniacal value-personas debates each architectural finding in a real codebase. A judge synthesizes the panel; an Opus report writer assembles the briefing. Anthropic prompt caching collapses the per-cell input cost so a full panel costs the same as one un-cached agent.
Every cell in a tribunal reads the same ~4–6 KB cached prefix: codebase summary, principle definitions, debate rules, decision-point evidence. The first cell writes the cache (paying 1.25× input price); cells 2–15 read it (paying 0.10× input price).
Per audit you'll typically see:
Cache savings: actual $0.420 · without cache $4.180 · 9.9× reduction
Per-cell cache hit rate (cells 0–5):
cell 0: 0.0% hit · read= 0t created= 4200t uncached= 510t $0.0079
cell 1: 89.2% hit · read= 4200t created= 0t uncached= 510t $0.0029
cell 2: 89.2% hit · read= 4200t created= 0t uncached= 510t $0.0029
...
Warm-cache cells (≥1): 89% hit rate across 14 cells
This is what makes a 15-agent panel economically viable. Without caching, every audit would cost ~10× more — and the panel size would have to shrink, which means losing the value diversity that drives the verdict quality.
- Layer 1 (deterministic): walks the repo, extracts decision points for ten structural principles — Martin's classic seven (cycles, stability, complexity, cohesion, reachability, layering, common-closure) plus stable-abstractions (P8 — I/A plane mis-placement), god classes/functions (P9 — size thresholds), and cross-file code duplication (P10 — jscpd). Off-the-shelf tools: vulture, radon, lizard, cppcheck, pydriller, jscpd.
- Layer 1.5 (cheap math): re-projects findings under the user's value weights.
- Layer 2 (LLM panel): 15 cells per finding. Each cell pairs two monomaniacal personas (Simplicity vs Velocity, Maintainer vs Shipper, etc.) who argue their value's reading of the evidence. A Sonnet judge renders one verdict per finding.
- Layer 3 (Opus briefing): one markdown report synthesizing every verdict into an audience-framed memo.
uv run forum audit ./your-repoOr with the live UI:
uvicorn server:app # http://localhost:8000