The shipped developer bundles

Four bundles ship with plnt and are what the parent picks from on a fresh install. All of them work on the session’s workspace with the built-in file tools; none needs a secret.

Bundle	Tools	Must call first	Config
code-reviewer	list_files, read_file, search	read_file	focus (bugs, security, style, all), max_findings, language_hint
test-writer	list_files, read_file, search, write_file, execute	read_file	framework (auto, pytest, unittest, jest, vitest, go), test_dir
repo-explainer	list_files, read_file, search	list_files	audience (new-contributor, reviewer, manager), max_words
changelog-writer	list_files, read_file, search, execute	none	style (keepachangelog, plain, release-notes), since

Must call first is the bundle’s require_tool: an answer given before that tool ran is withheld. A reviewer that never read a file does not get to report findings.

In read-only mode

The public playground and any server with PLNT_READ_ONLY=1 strip write_file and execute before the agent sees its tools. test-writer then returns the test file content in its answer and says it was not written or run; changelog-writer writes from the current code instead of git history and says so.

Installing and tuning

plnt install code-reviewer --tenant acme --config focus=security --config max_findings=5
plnt install test-writer --tenant acme --config framework=pytest --config test_dir=tests

The console’s Agents tab builds a settings form from each bundle’s config_schema.json.

What the parent is told

For each installed bundle the parent sees its name, description and tools. The descriptions are written for that: they say what the bundle is for, so the parent can split a task by concern (“audit” to the reviewer, “tests” to the test writer, “explain” to the explainer).

Your own

Copy one of them as a starting point:

cp -r "$(python -c 'import plnt.bundles.catalog as c; print(c.resolve("code-reviewer").path)')" ./my-reviewer

then edit skill.toml, prompt.md and config_schema.json. See Write a bundle.