DEV Community: Toruk Makto

Reading your AI coding logs: cache hits, retry loops, and other signals

Toruk Makto — Thu, 16 Apr 2026 09:58:49 +0000

Last week I checked my AI coding spend and it was higher than my AWS bill. I'm paying for Claude Code, Codex, Cursor, the occasional Opus burst, and I had no visibility into where any of it went. Just a number going up.

Turns out every AI coding tool already writes session data to disk. Claude Code drops JSONL into ~/.claude/projects/. Codex writes to ~/.codex/sessions/YYYY/MM/DD/. Cursor uses a SQLite database. OpenCode uses SQLite. Pi uses JSONL. All of it is sitting there waiting to be read.

I started reading mine and the patterns are obvious once you look.

What the data shows

This is one week of my actual AI coding usage:

A few things jumped out immediately.

Cache hit rate matters more than I thought. Claude prices cache reads at 1/10th the cost of fresh input. Opus came in at 98.8% cache hits, which sounds great until I noticed Sonnet 4.6 was at 77.1%. That gap is real money. If your system prompt or the first few files in context are unstable, you're paying full price for the same tokens every turn.

Tool counts tell you the agent's mood. 2,126 Bash calls, 990 Reads, 742 Edits in a week. The Read:Edit ratio is roughly 1.3, which is fine. If Read had been 4x higher, I'd know the agent was spelunking instead of executing.

One-shot rate is brutal honesty. Coding shows 88% one shot. The other 12% needed retries (Edit → Bash → Edit). That's where time and tokens leak silently.

Model mix reveals overspending. Opus 4.6 cost $1219 this week. Sonnet 4.6 cost $38. Some of those Opus turns were small Q&A that Sonnet would have handled fine. I haven't run the experiment of routing them yet, but the gap suggests there's real money on the table.

Patterns worth watching for

Signal	What it usually means
Cache hit < 80%	System prompt or context unstable, caching not configured
Lots of Read calls per session	Agent re-reading files, missing context
Low one-shot rate (Coding < 30%)	Retry loops, agent struggling with edits
Opus dominating cost on small turns	Overpowered for the task
`dispatch_agent` heavy	Sub-agent fan-out, expected or excessive
No MCP usage	Either you don't use MCP, or your config is broken
Bash dominated by `git status`, `ls`	Agent exploring instead of executing
Conversation category dominant	Agent talking instead of doing

These aren't verdicts, just starting points. A 60% cache hit on a one-off experiment is fine. A persistent 60% across weeks is a config issue.

How I'm reading this data

There's a tool called codeburn that reads all the session formats and renders this dashboard in your terminal. It supports Claude Code, Codex, Cursor, OpenCode, and Pi. No proxy, no API keys just reads the local files.

npx codeburn report --period week

Repo: https://github.com/AgentSeal/codeburn

Open source, MIT.

Why this matters

We obsess over model choice and pricing tier. We argue about Opus vs Sonnet vs GLM. The discussion online is almost entirely about which model to use, never about what your agent is actually doing once it's running.

The session files have the answer. Every retry, every redundant Read, every cache miss, every misrouted model,it's all there. Looking at it once a week takes ten minuts and tells you more about your spend than any pricing comparison.

Try reading your own sessions for a week. Even if you don't use any tool, just cat a few JSONL files and look at the usage blocks. You'll spot at least one pattern you didn't expect.

Where do your AI coding tokens actually go?

Toruk Makto — Thu, 16 Apr 2026 09:49:05 +0000