Analytics & Cost Tracking¶

CodePlane tracks every token, tool call, and dollar across all jobs — giving you fleet-wide visibility into what your coding agents cost and how they perform. All analytics are computed from the normalized provenance data layer — the same SQLite database you can query directly. Open the analytics dashboard with Alt+A.

Scorecard¶

The scorecard is the top-level summary. It shows per-SDK budget totals, job activity breakdown, and daily cost trends over a configurable period (7–365 days).

Budget by SDK — Total spend for each SDK (Copilot, Claude, etc.) with cost trends
Activity breakdown — Jobs by resolution: running, merged, PR created, discarded, failed, cancelled
Copilot quota — If you use Copilot, the scorecard tracks premium request consumption and alerts when quota exceeds 80%
Daily cost trend — Area chart showing spend over time

Understanding costs

For subscription plans (like Claude Max or Copilot Business), CodePlane shows what the same usage would cost at API rates. This gives you a consistent cost metric for comparing models and optimizing agent behavior, even when you're on a flat-rate plan.

Model Comparison¶

Compare models head-to-head on cost, speed, and outcomes.

Metric	Description
Avg Cost	Average USD per job for each model
Avg Duration	Average job runtime
Cost/min	Spend efficiency — lower is better
Cost/turn	How much each agent turn costs on average
Resolution rates	Per-model breakdown of merged / PR'd / discarded / failed

Filter by repository to compare model performance on specific codebases.

Repository Breakdown¶

See which repos drive the most spend and activity.

Cost, job count, and token totals per repository
Tool calls and average job duration
Premium request consumption (Copilot)

Tool Health¶

Monitor the reliability and latency of every tool your agents use.

Call counts — How often each tool is invoked
Failure rate — Percentage of calls that errored (flagged when ≥20%, critical at ≥50%)
Latency — Average, p50, p95, and p99 durations
Tool categories — file_write, file_read, file_search, git, shell, browser, agent, system

Cost Drivers¶

Identify which jobs, models, and repos contribute the most to your spend.

Activity Breakdown¶

Cost attribution is classified by activity dimension — what the agent was doing when it spent tokens:

Activity	Description
Implementation	File edits and code generation
Verification	Test execution and validation
Investigation	Reading, searching, browsing code
Git ops	Commits, pushes, diffs
Setup	Package installs, environment configuration
Delegation	Sub-agent calls
Reasoning	Pure thinking without tool use
Communication	User-facing responses

Each turn is classified by an LLM that observes the agent's intent, providing accurate cost allocation even when a single turn mixes multiple tools.

Latency Breakdown¶

Parallel to the cost breakdown, the latency view shows wall-clock time attribution by activity dimension — where time is actually spent during job execution. Includes percentile metrics (p50, p95) per dimension.

Tool Mix¶

A percentage breakdown showing which tools the agent uses most (by call count and cost contribution), grouped by activity category.

Per-File Cost Attribution¶

Cost is also attributed at the file level — see which files drove the most agent spend. This helps identify files that are expensive to modify (complex logic, many callers, poor test coverage) and files the agent rereads excessively.

Token & Cache Metrics¶

Every job tracks token usage in detail:

Input tokens and output tokens (separately)
Cache read tokens and cache write tokens (prompt caching)
Cache hit rate — percentage of input tokens served from cache
Per-model and per-repo token aggregations

Daily Spend Limit¶

Set a personal daily budget in your config:

telemetry:
  daily_spend_limit_usd: 10.00

When configured, the Budget card shows a progress bar for today's spend vs. your limit. Warnings appear when you cross 80%.

Observations (Smart Alerts)¶

CodePlane runs statistical analysis across your jobs to surface actionable cost observations:

Detector	What it catches
File rereads	Files read excessively across jobs (≥10 reads, ≥3 jobs, >10KB total)
Tool failures	Tools with ≥20% failure rate across ≥10 calls
Turn escalation	Jobs where 2nd-half cost is ≥2× 1st-half and ≥$0.50
Retry waste	Tools retried ≥10% of the time
Compaction storms	Jobs with ≥5 context compactions
Cache regression	Cache hit rate drops ≥15pp vs. prior week

Observations appear as alert banners at the top of the analytics page. Dismiss them individually once reviewed.

CSV Export¶

Tables in the analytics dashboard (Jobs, Model Comparison) include a CSV button that exports the visible data for use in spreadsheets or external analysis tools.

Hub Architecture (Future)

CodePlane is a personal-first tool. Each instance runs locally with its own SQLite database. For teams that want aggregate visibility, a future CodePlane Hub will accept telemetry pushes from personal instances:

Each instance has an auto-generated instance_id (in telemetry config)
The JobTelemetryReport schema defines the per-job payload
Instances push completed-job summaries to the Hub endpoint
The Hub aggregates fleet-wide analytics without accessing source code

This is not yet implemented — the schema and instance ID are in place as foundations.