analysis-patterns

Analysis Patterns — Reusable Heuristics

Standardized analysis methods used by multiple agents.

Tool Sequence Patterns

Productive Sequences (high correlation with quality score ≥4)

Pattern	Description	Expected Token Cost
Read→Edit→Bash(pass)	Read file, edit, verify	~1500 tokens
Glob→Read→Edit	Find file, read, edit	~1800 tokens
Read→Task(sub)→Edit	Read, delegate, apply	~2500 tokens

Anti-Patterns (high correlation with quality score ≤3)

Pattern	Description	Token Waste
Read→Read→Read(same)	Repeated reads of same file	~1000 tokens/repeat
Bash(fail)→Bash(fail)→Edit	Build-before-edit	~800 tokens wasted
Edit→Bash(fail)→Edit→Bash(fail)	Edit-test loop without reading error	~1500 tokens/cycle

Statistical Methods

Z-Score Calculation

z = (x - μ) / σ
Threshold: |z| > 2.0 → anomaly
Minimum data points: 5 (for reliable σ)

Moving Average

MA(n) = sum(last_n_values) / n
Window: 5 sessions (default)
Alert if: current > MA × 1.5

Efficiency Ratio

efficiency = productive_calls / total_calls
productive_call = call that contributed to final output (no retry, no revert)
Target: ≥ 0.85

Token Estimation Baselines

Expected token usage per tool (for waste detection):

Tool	Expected Range	Flag If
Read	200-1500	>2000 (large file, use offset)
Edit	100-500	>800 (complex edit, consider split)
Write	200-2000	>3000 (generated file too large)
Bash	50-500	>1000 (verbose output, use --quiet)
Glob	50-200	>500 (too many matches)
Grep	50-300	>800 (broad search pattern)
Task	500-5000	>8000 (sub-agent context explosion)

Plugin Phase Patterns

Phase detection and per-plugin baselines are loaded dynamically from plugin profiles: ~/.claude/plugin-introspector/plugin-profiles/{plugin}/profile.json

When baselines unavailable (< 5 sessions), use universal defaults:

• Entry phase: 5-15% of session tokens
• Middle phases: 35-60% combined
• Exit phase: 15-30% of session tokens

Phase-Generic Anti-Patterns (PG-001~PG-007) are defined in improvement-pipeline.md Layer 2.

Improvement Signal Extraction

From Quality Evaluator

Field names match quality-evaluator.md improvement_signals[] output:

IF dimension.score ≤ 3:
  signal = {
    dimension: dimension.name,
    score: dimension.score,
    score_gap: 5 - dimension.score,
    what: "description of the issue",
    root_cause: "underlying cause",
    quantified_impact: {tokens, cost_usd, percentage_of_session},
    trace_evidence: ["trace_id: description", ...],
    suggested_change: {target_file, change_type, description},
    priority: score_gap × dimension.weight
  }

From Token Optimizer

IF waste_source.tokens_wasted > total_tokens × 0.05:
  signal = {
    type: waste_source.type,
    impact: waste_source.tokens_wasted,
    priority: impact / total_tokens
  }

From Anomaly Detector

IF alert.severity == "HIGH":
  signal = {
    type: alert.type,
    recurrence: count(similar_alerts_in_history),
    priority: recurrence > 2 ? "CRITICAL" : "HIGH"
  }