Back to Overview

Sidechain Rate

Agent Behavior

· 47 PRs with data in past 30d (118 total)

Count
47
Average
17%
↑ 2% vs prior 30d
P10
6%
↑ 1% vs prior 30d
P50
18%
↑ 2% vs prior 30d
P90
26%
↑ 1% vs prior 30d

Trend

Distribution

0–10%
7
10–20%
21
20–30%
19

Notable PRs

Highest
#155Migrate to new logging framework28%
#125Add Prometheus metrics endpoint28%
#117Implement soft delete for projects27%
Lowest
#196Add custom domain support3%
#195Fix webhook signature verification3%
#142Fix dropdown z-index stacking3%

About This Metric

Sidechain Rate

What It Measures

The fraction of messages in a session that occurred on sidechain branches — alternate reasoning paths the model explored and then abandoned. Sidechain messages represent work the model did that didn't contribute to the final output.

Why It Matters

When Claude Code backtracks, it creates sidechain branches — invisible to the user but consuming tokens and time. A high sidechain rate means the model is frequently going down wrong paths before finding the right approach. This is a direct measure of wasted effort.

Tracking sidechain rate helps identify tasks or prompting patterns that confuse the model. If certain types of requests consistently produce high sidechain rates, that's a signal to improve prompt clarity, provide better context via CLAUDE.md, or break the task into smaller pieces.

How It's Calculated

sidechain_rate = sidechain_messages / (human_messages + assistant_messages)

Summed across all sessions correlated to the PR. Returns a value between 0.0 and 1.0. Returns null if there are no messages.

Data Sources Required

  • Claude Code session data — Message-level isSidechain flag from session JSONL files.