Observer

A 12-module analytics engine for AI-assisted development sessions. Tracks tokens, costs, tool calls, timing, and patterns -- then runs intelligence analysis to surface actionable insights.

What the Observer Does#

The observer collects metrics during every AI assistant session and runs them through 12 intelligence modules. It answers questions like:

• How efficiently am I using the AI assistant?
• Am I exhibiting anti-patterns (retry storms, context bloat, edit churn)?
• What is my projected session cost?
• Is the context window filling up? When will compaction be needed?

All analysis is local. No data leaves your machine.

Core Tracking#

Every session automatically tracks:

• Tokens -- Input, output, and total token consumption per turn
• Costs -- Estimated spend based on the model in use (Opus, Sonnet, Haiku)
• Tool calls -- Count, types, success/failure ratio, duration
• Timing -- Session duration, time per turn, idle gaps
• Patterns -- Repeated file edits, search patterns, tool sequences

Intelligence Modules#

The bootspring observe intelligence command runs all 12 modules:

Efficiency Scorer#

6-dimension weighted geometric mean producing a 0-100 composite score with a letter grade (A-F). Dimensions: task completion rate, tool accuracy, context utilization, cost efficiency, time management, code quality.

Session Fingerprint#

Maps sessions to a 16-dimensional feature vector and classifies them into 8 types (exploration, implementation, debugging, refactoring, testing, review, planning, mixed) using nearest-centroid classification.

Context Pressure#

Tracks context window fill rate over time. Uses EWMA (exponentially weighted moving average) to predict when compaction will be needed. Four pressure levels: low, moderate, high, critical.

Pattern Detector#

Detects 8 anti-patterns:

• Thrashing -- Repeatedly switching between unrelated tasks
• Blind search -- Searching without narrowing scope
• Retry storm -- Repeating the same failing operation
• Context bloat -- Loading unnecessary files into context
• Tool misuse -- Using the wrong tool for a task
• Prompt starvation -- Not providing enough context in prompts
• Subagent sprawl -- Spawning too many parallel agents
• Edit churn -- Editing the same file many times without converging

Drift Detector#

Uses binary segmentation to find change points in session behavior. Calculates Jensen-Shannon divergence between phases. Produces phase summaries showing how the session evolved.

Predictive Cost#

EWMA burn rate estimation with linear regression extrapolation. Runs Monte Carlo simulation to produce a 95% confidence interval for total session cost. Predicts budget exhaustion time.

Anomaly Detector#

Combines four statistical methods to flag unusual events: modified Z-score with MAD, CUSUM (cumulative sum), IQR fencing, and Grubbs' test. Detects 6 anomaly types.

Time Series#

Simple moving average, trend detection, change point identification, and learning curve fitting across 7 metrics. Shows whether performance is improving, stable, or declining.

Session Comparator#

Compares two sessions metric-by-metric with significance testing. Produces a weighted Jaccard similarity score. Useful for A/B testing different approaches.

Cost Optimizer#

Analyzes token waste, compares costs across models (Opus vs Sonnet vs Haiku), and projects monthly spend. Recommends model switches for exploratory vs focused work.

Smart Alerts#

Dynamic threshold alerting with suppression, escalation, and rate limiting. Prevents alert fatigue while ensuring critical issues surface.

Intelligence Hub#

Orchestrates all modules, isolates errors, synthesizes cross-module insights, and produces a composite health score.

Three Modes#

CLI Commands#

Further Reading#

• CLI reference for observe commands
• Session Intelligence -- the unified telemetry layer
• Hooks -- the event system that feeds data to the observer