Session Trends

SkillDocs & knowledge

Analyze trends across session metrics. Computes windowed aggregates, deltas, and compares against MEMORY.md findings. Use periodically for progress tracking.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the Session Trends skill

What this skill tells your AI

The instructions your AI receives, as published by oliver-kriska/claude-elixir-phoenix in .claude/skills/session-trends/SKILL.md and read by ahel’s review.

Analyze trends from the metrics ledger. Computes windowed aggregates, fingerprint distributions, and compares against MEMORY.md baselines.

Requirements

Requires .claude/session-metrics/metrics.jsonl from /session-scan.

Usage

/session-trends                          # All windows (7d, 30d, all)
/session-trends --window 30d             # Specific window only
/session-trends --project enaia          # Filter by project
/session-trends --compare MEMORY.md      # Compare against memory baseline
/session-trends --html out.html          # Write HTML report with ASCII bars

For pure context-window stats (max prompt tokens, ctx %, compaction rate) across raw Claude Code JSONL files, see the --scan-jsonl mode of compute-metrics.py (inspired by badlogic / earendil-works/pi).

Pipeline

Step 1: Parse Arguments

Extract from $ARGUMENTS:

  • --window WINDOW: Time window — 7d, 30d, or all (default: show all three)
  • --project NAME: Filter metrics by project name
  • --compare PATH: Path to MEMORY.md for baseline comparison (default: auto-detect from .claude/ project memory)

Step 2: Read Metrics Ledger

Read .claude/session-metrics/metrics.jsonl.

If empty or missing:

No metrics found. Run /session-scan first.

If --project specified, filter entries by project field.

Step 3: Compute Trends via Python

python3 .claude/skills/session-scan/references/compute-metrics.py \
  --trends .claude/session-metrics/metrics.jsonl \
  --memory {MEMORY_PATH}

Capture the JSON output.

Step 4: Display Trend Report

Format the JSON output as a readable report:

Overview
Total sessions: {N} ({backfilled} backfilled from v1)
Date range: {earliest} to {latest}
Window Comparison
| Metric                  | 7 days | 30 days | All time |
|-------------------------|--------|---------|----------|
| Sessions                | 12     | 45      | 165      |
| Avg friction            | 0.28   | 0.24    | 0.22     |
| Max friction            | 0.72   | 0.72    | 0.89     |
| Avg opportunity         | 0.35   | 0.30    | 0.28     |
| Tier 2 eligible         | 40%    | 33%     | 30%      |
| Plugin adoption         | 12%    | 10%     | 8%       |
Fingerprint Distribution
| Type          | 7d  | 30d | All  |
|---------------|-----|-----|------|
| bug-fix       | 4   | 15  | 52   |
| feature       | 3   | 12  | 48   |
| exploration   | 2   | 8   | 30   |
| maintenance   | 1   | 5   | 18   |
| review        | 1   | 3   | 10   |
| refactoring   | 1   | 2   | 7    |
MEMORY.md Comparison (if --compare)

Compare measured values against MEMORY.md claims:

| MEMORY.md Claim              | Measured    | Match? |
|------------------------------|-------------|--------|
| Plugin adoption: 8-12%       | 10.2%       | Yes    |
| Minimal friction in 40+ of 74| 68% smooth  | Yes    |

Step 5: Write trends.json

Write computed trends to .claude/session-metrics/trends.json.

Step 6: Suggest Actions

Based on trends:

  • If friction is increasing: "Friction trending up — run /session-deep-dive --from-scan to investigate"
  • If plugin adoption is growing: "Plugin adoption growing — check which commands drive value"
  • If many Tier 2 eligible: "{N} sessions need deep analysis"

Output Files

FilePurpose
.claude/session-metrics/trends.jsonComputed trend data

Common Queries

See references/trend-queries.md for interpreting specific trend patterns.

Iron Laws

  1. ALWAYS use Python for computation — no manual aggregation
  2. NEVER modify metrics.jsonl — read-only for trends
  3. ALWAYS show window comparison — single numbers lack context

Acknowledgements

The HTML report layout (preformatted text + ASCII bar charts via /) and per-model + threshold-bucket breakdown (>=80%, >=90%, >=100%, compaction_rate) were borrowed from badlogic / earendil-works/pi session-context-stats.mjs. Our pipeline's qualitative metrics (friction, fingerprint, plugin opportunity, skill effectiveness) are additive on top.

Signals

GitHub stars
543
Forks
38
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
session-trends
Source
github.com/oliver-kriska/claude-elixir-phoenix