Performance Metrics

Astrakion operates under a Kanban workflow where success is defined by stable flow, predictable delivery, small batch sizes, minimal waiting, and consistent cycle times.


Kanban Principles

  • Flow > Speed — Predictability over raw execution speed
  • Small Work Items — Issues fit into Kade's 5-15 minute window
  • Limit WIP — One issue moves through the pipeline at a time
  • Make Delays Visible — Blocked time identifies systemic delays
  • Metrics Guide Refinement Only — Metrics never cause rework

Core Metrics

The following metrics are recorded for every code-impacting issue. Analysis-only issues do not generate metrics.

Lead Time

Time from human approval → Astra submitting a PR into <HOD_BRANCH>.

Represents end-to-end delivery from the human's perspective.

Cycle Time

Time from Astra assigning to Kade → Orion merging into astrakion/develop.

Shows internal AI execution efficiency.

Implementation Time

Time from Kade beginning → Kade opening the PR.

Ensures tasks stay within 5-15 minute batch size.

Estimate Accuracy

Difference between Kade's estimated time and actual implementation time:

  • Accurate: ±2 minutes
  • Moderate variance: ±3-5 minutes
  • High variance: 6+ minutes

Helps detect unclear requirements or hidden complexity.

Flow Efficiency

Ratio of active work time to total cycle time:

Flow Efficiency = Active Work Time / Total Cycle Time

Shows how much time is spent working vs. waiting.

Blocked Time

Time an issue cannot progress due to:

  • Waiting for human clarification
  • Human-origin PR blocking the AI lane
  • LLM provider errors
  • Exhausted Astrakion Tokens
  • GitHub outages

Analysis-only issues never contribute to Blocked Time.

Throughput

Number of code-impacting issues completed per day or week.

Measures long-term delivery stability.

Work Type Distribution

Percentage breakdown by category: bug fix, feature, enhancement, clarification, QA-raised improvement, human-directed update.

Cost Per Issue

Effective cost to complete one code-impacting issue based on tier, included capacity, and tokens consumed.

Analysis-only issues have no cost.


What Astra Learns

Metrics inform future refinement:

  • Oversized Issues — Long cycle time → future issues should be split
  • Unclear Requirements — High variance → improve refinement
  • Risky Modules — Repeatedly slow areas tracked for caution
  • Human Bottlenecks — High blocked time from clarifications
  • Predictability Trends — Stable throughput indicates healthy flow

Metrics never affect current work — only future tasks.


Safety Guarantees

Performance metrics never trigger:

  • Rework or retries
  • Issue reopening
  • PR modification
  • Additional test passes
  • Scope expansion
  • Automatic corrections to rejected PRs

Ownership Summary

Metric Owner
Lead TimeAstra
Cycle TimeAstra
Implementation TimeKade → Astra
Estimate AccuracyAstra
Flow EfficiencyAstra
Blocked TimeAstra
ThroughputAstra
Work Type DistributionAstra
QA / Risk InsightsOrion
Cost Per IssueAstra