After digging through local session logs, the recurring issue was not just model pricing. A lot of the waste came from context bloat: generated artifacts, logs, build output, oversized CLAUDE.md / AGENTS.md files, repeated tool output, broad repo exploration, stale session state, and command loops.

PrismoDev is a local CLI that scans a repo plus local Claude Code / Codex logs and tries to explain where token/context waste is coming from. It does not require API keys or login, and nothing leaves the machine.

A few commands:

```bash npx getprismo doctor ```

Scans the repo, flags missing `.claudeignore` / `.cursorignore`, oversized instruction files, generated artifacts exposed to agents, and creates compact `.prismo/` context summaries.

```bash npx getprismo watch ```

Monitors live context pressure during a coding session and warns about repeated file reads, artifact leaks, oversized tool output, and possible loops.

```bash npx getprismo cc timeline ```

Builds a postmortem timeline showing where context spikes, repeated reads, command loops, or generated artifacts appeared in a session.

There are also scoped context policies:

```bash npx getprismo firewall auth-bug ```

which creates a task-specific context boundary before starting work.

The main idea is to make AI coding sessions easier to reason about before the bill surprises you: what got loaded, what repeated, what could have been avoided, and what should be ignored next time.

Repo: https://github.com/shanirsh/prismodev

I’d be especially interested in feedback from people using Claude Code, Codex, Cursor, or similar tools on larger repos. I’m trying to tune the false positives and find more real-world waste patterns.