I released v4 of cc-thingz.
It is my open-source toolbox for AI coding agents: skills, agents, hooks, and safety rails for real coding workflows.
It works across:
- Claude Code
- OpenAI Codex CLI
- Google Gemini CLI
- Pi
Tachles: this is not another pile of prompts.
It is the same workflow logic, authored once, then compiled into each tool’s native shape.
Why v4 exists
Before v4, the project was still too Claude-shaped.
Other tools were supported, but the setup had too much duplication and too much:
why is this different here?
Classic balagan.
Now everything is authored once under src/ and compiled into platform-specific output.
What that buys you
- shared workflow logic lives once
- each tool gets instructions in its own shape
- hooks and manifests are generated, not hand-juggled
- tests catch drift before users do
Not glamorous. Very useful.
The infra got more serious too
One thing I ended up caring about more than expected: the prompts are not managed like random markdown blobs.
Skills live in canonical SKILL.md files, then get small per-tool overlays where behavior really differs. The build strips vendor-specific bits out of the shared base, validates that the base stays vendor-neutral, and snapshot-tests generated outputs so drift gets caught early.
There are also paid skill eval fixtures for regression testing behavior, because “seems fine” is not much of a test strategy when you are shipping instructions to models.
That part is not the headline, but it is why the portability claim is real instead of decorative.
The part I care about most: multi-agent support
The bigger value is multi-agent work that does not depend on one vendor UI.
cc-thingz now ships a larger shared agent set for:
- review
- implementation
- testing
- docs
- language work
- infra
- planning
- exploration
Claude Code and Pi can both use that shared set.
In Pi, this works through the @tintinweb/pi-subagents runtime instead of some separate Pi-only agent universe.
Pi also gets four pipeline agents:
scoutplannerreviewerworker
Those are there to decompose work instead of stuffing everything into one heroic context window and pretending it is an engineering team.
Small specialized agents. Bounded jobs. Clear handoffs.
Less magic. More plumbing. Plumbing wins.
Hooks got properly portable too
Hooks are part of the point.
v4 keeps pushing safety into automatic checks:
- linting
- test running
- git guardrails
- session context
- protected-path handling
And this bit matters more than it sounds:
Pi now bridges its own lifecycle and tool events to the same hook scripts and exit-code semantics.
That means existing hook logic can run there too, and user-defined hooks can be layered on top instead of rebuilding the whole thing from scratch.
Same guardrails. More than one agent. Much better.
Recent v4 work also made the file-protection hook understand Codex patch-based edits, not just simple file writes.
If an agent can edit multiple files in one patch, your guardrails need to understand patches.
Otherwise it is not a guardrail.
It is decoration.
What it is for
Use it if you work with coding agents and want reusable workflows for:
- code review and fixes
- Go, Python, TypeScript, web, infra
- docs, tests, research, worktrees
- spec-driven development
- multi-agent planning and review
- boring safety checks before expensive mistakes
Repo
It is opinionated because neutral tools become mush.
Built from daily use, not from a weekend prompt dump with a logo.
Top comments (0)