codebase-analysis

Codebase Analysis for Nerd

Parameter Categories (by impact)

High: Search/ranking weights, resolution thresholds, AI prompt efficiency, token budgets Medium: Temporal decay rates, cache TTLs, batch/concurrency limits, scoring formulas Lower: Animation timing, layout defaults, retry/backoff parameters

Scan Patterns

# Constants and config values
grep -rn "const\|DEFAULT_\|THRESHOLD\|LIMIT\|MAX_\|MIN_\|TIMEOUT\|BATCH\|WEIGHT\|BOOST\|DECAY"

# Float comparisons (thresholds)
grep -rn ">= 0\.\|<= 0\.\|> [0-9]\.\|< [0-9]\."

# Prompt templates
grep -rn "system_prompt\|systemPrompt\|PROMPT\|template\|instructions"

# Tuning TODOs
grep -rn "TODO.*tun\|FIXME.*threshold\|TODO.*calibrat\|TODO.*magic"

Skip: math constants, UI colors/padding, protocol values (HTTP codes), test fixtures, values with citations, parameters in non-executable files (markdown, docs) unless flagged as analytical.

Prioritize by: measurability first (can a shell command output a number?), then frequency (per-request > per-session), sweep feasibility (no side effects), data availability (feedback/ground truth exists).

Measurability Classification

• Experimentable: parameter is in executable code, changing it produces a measurable effect on an automated metric → experiment_type: "parameter_sweep" or "comparison" or "ablation"
• Analytical: parameter is in non-executable files (markdown, config comments, agent prompts) or requires human judgment to evaluate → experiment_type: "analytical" — suitable for /nerd batch analysis but NOT for /nerd-loop