noisy-tools-in-subagent

noisy-tools-in-subagent

Forces noisy commands — builds, tests, linters, static analysis, anything that produces a wall of output — to run inside a Sonnet subagent instead of the main agent's context window. Main-agent tokens are for work that needs them, not for 2,000 lines of Maven output.

Installation

Prereq: install bash-classify, used by the hook to parse Bash commands:

uv tool install bash-classify

Then add the marketplace and install:

claude plugin marketplace add fprochazka/claude-code-plugins --scope user
claude plugin install noisy-tools-in-subagent@fprochazka-claude-code-plugins --scope user

If bash-classify isn't on your PATH, the hook refuses to proceed and tells the agent to install it.

How it works

Two pieces:

1. A PreToolUse hook on Bash that parses each command with bash-classify (walking pipes, chains, and subshells) and matches every parsed command against a regex whitelist of noisy build/test/lint tools. In the main agent context, a match gets denied with a short message instructing the agent to delegate to the subagent. In any subagent context, the hook passes through — other subagents aren't affected, and the noisy-runner can freely run what it was asked to run.

2. A noisy-runner subagent (Sonnet, strictly scoped to Bash and Read) that runs the commands it is given, wraps them in tee to capture full logs to /tmp, reads any files the output references to interpret errors, and reports back a concise summary plus log paths. It does not explore, fix, or modify anything — its only job is run, read, interpret, report.

Net effect: when the main agent wants ./mvnw test, it gets redirected into a "3 tests failed at X:42, Y:87, Z:12 because …, full logs at /tmp/…" summary instead of a flood of output.

The whitelist and its tuning live in hooks/enforce-subagent.py. Small introspection commands (mvn help:*, gradle tasks, mvn dependency:tree) pass through to the main agent — only the heavy lifecycle phases are blocked.

Using the subagent directly

The noisy-runner subagent isn't limited to the hook's whitelist. Invoke it via the Task tool any time you want a command run and interpreted rather than transcribed into your context:

subagent_type: "noisy-tools-in-subagent:noisy-runner"
prompt: run `./mvnw test -pl foo` and `./gradlew :bar:check`, tell me what broke

You can pass multiple commands in one invocation — the subagent runs them sequentially.

Known Claude Code UI quirk

When the hook denies a command, Claude Code's terminal UI may render the rejection message 2–3 times in the same tool-call block (an empty "blocking error" header, a red Error: card, sometimes a side-panel overlay). This is a Claude Code rendering quirk — upstream issues anthropics/claude-code#34713 and #21504. The model itself only sees the rejection reason once in its context, so it's visual clutter, not token waste.

Upgrading

claude plugin marketplace update fprochazka-claude-code-plugins
claude plugin update noisy-tools-in-subagent@fprochazka-claude-code-plugins

License

MIT