Bash command wrappers / prefixes (sudo, npx, rtk, …) are misclassified

[athoune]

Bash command wrappers / prefixes (sudo, npx, rtk, …) are misclassified

Problem

When an AI agent prefixes its bash commands with a wrapper tool, all those commands get lumped under the wrapper name in CodeBurn's bashBreakdown. For example, an agent that runs every command through a prefix like rtk (or sudo, npx, etc.) produces entries like this:

Raw command	Currently classified as	Should be
rtk git push	rtk	git
rtk npm install	rtk	npm
sudo apt update	sudo	apt
npx vitest --run	npx	vitest

This makes the bash command breakdown and the optimize detectors (which look at bash output waste) much less useful when a prefix is in use, because the actual tool being invoked is invisible.

The cause is in src/bash-utils.ts — extractBashCommands takes basename() of the first non-env-var token without considering that the token might be a known wrapper that delegates to a real command.

Proposed solution

Add a set of known command wrappers in bash-utils.ts and skip past them during extraction:

const COMMAND_PREFIXES = new Set([
  'sudo', 'doas',       // privilege elevation
  'npx', 'bunx', 'dx',  // package runners
  'time',               // timing wrapper
  'nice', 'nohup', 'stdbuf', // process control
  'rtk',                // user-land command prefix
])

The skip only activates when there is a next token (a bare rtk or sudo without a following command is kept as-is).

Scope

The change is localised to bash-utils.ts — the same function is called by every provider (40+ call sites across all parsers), so the improvement applies universally.

Tests

• rtk git status → ['git']
• sudo npm install → ['npm']
• npx vitest --run → ['vitest']
• DEBUG=1 rtk git status → ['git'] (env var + prefix combined)
• sudo npx vitest --run → ['vitest'] (nested prefixes)
• rtk git add . && rtk git commit -m "msg" → ['git', 'git'] (chained)
• rtk alone → ['rtk'] (standalone prefix kept)
• Existing tests unchanged: npm install && git push → ['npm', 'git']

Prior art

The function already skips leading VAR=value assignments and filters out cd / true / false. The prefix skip is a natural extension of the same pattern.

I have a working branch with the changes and all tests passing. Happy to open a PR if the approach sounds right.

Additional context

• CodeBurn version: 0.9.15
• This affects all providers that go through extractBashCommands (Claude, Cursor, OpenCode, Gemini, Codex, Copilot, and 15+ more)
• The rtk prefix is used by some agent configurations; sudo, npx, bunx etc. affect anyone whose agent uses privilege elevation or package runners

[iamtoruk]

Thanks for the clear writeup and the proposed approach, this was spot on. Fixed in #658.

Implemented the wrapper skip in extractBashCommands for sudo, doas, npx, bunx, time, nice, nohup, stdbuf, and rtk, interleaved with the existing VAR=value skip so mixed forms like sudo NODE_ENV=prod node server.js resolve to the real tool.

Two deviations from the proposed set, both to avoid emitting misleading keys:

• Dropped dx from the list. It's the Dioxus CLI binary (dx serve, dx build), a real tool people run directly, so skipping it would turn dx serve into a bare serve key.
• A wrapper followed by a flag or a quoted token is kept as-is (nice -n 10 git stays nice, npx "@angular/cli" stays npx), since we can't reliably resolve the real command in those forms without knowing each flag's arity, and guessing would produce garbage keys.

All your listed test cases pass, plus added coverage for the env+wrapper interleave and the quote guard. Closing this as resolved by #658, thanks again.