ai-agents-metrics

Analyze your AI agent work history. Track spending. Optimize your workflow.

AI is writing more of your code. You still don't know:

• How many attempts each task actually takes
• Where the process breaks down and why
• Whether your workflow is getting faster or generating more rework

ai-agents-metrics extracts these signals from your existing Claude Code or Codex history. Point it at your history files and see what's happening: retry pressure, token cost, and session timeline.

Running this on 6 months of Claude Code + Codex history (3.85B tokens, 160 threads) surfaced:

• 100% of Claude "retries" are subagent spawns, not user retries — attempt_count > 1 is structural, not a failure signal (F-001)
• Subagent delegation halves main-session tokens within-thread — median 2.05× compression, p = 0.000456 (F-007)
• Per-skill compression ranking — Explore 2.63×, code-reviewer 3.25×, commit 0.72× (F-008)

Full index: docs/findings/. N=1 developer; the mechanisms generalize because they come from the tools, not the data.

---

Quick start

pipx install ai-agents-metrics

ai-agents-metrics history-update     # reads ~/.codex + ~/.claude by default
ai-agents-metrics show               # retry pressure, cost, session timeline
ai-agents-metrics render-html        # interactive HTML report

Non-default history paths and the full command list: CLI reference.

---

What you get

• History extraction — retry pressure, token cost, model usage from existing session files. No setup.
• HTML report — one self-contained file, summary strip + 5 trend charts, opens in any browser.

Not a benchmark, not an eval framework, not a model comparison tool. It is a local analysis tool for real engineering work done with AI.

---

Privacy

All data stays local. Writes only to:

• .ai-agents-metrics/warehouse.db — local SQLite warehouse used by the history pipeline No data is sent to any remote service.

---

Links

• Repository · Changelog · Contributing · Security
• CLI reference · Findings · Architecture