Convert and Enhance Video Skill Extension

Summary

@skills/creative/manim-video/ lets load this skill into pi as an extension in ~/.pi/

Readextensions.md

ReadREADME.md

ReadSKILL.md

Readskills.ts

Readindex.ts

Readskills.md

lets not use @skills/creative/manim-video/scripts/setup.sh directly. the requirements are installed. lets use a lighter version that will use uv for the python and the manim executatble and then checks for the other two, but just asks the user to install manually instead of just installing shit

we'll model this as a simple pi extension that works with /manim

given these instructions, can you port everything under @skills/creative/manim-video/ to a new pi extension? do it in a subrepo in ~/dev/ use pnpm dlx gh-axi to create a new repo and then push to it call it pi-manim-video install it via pi install git:[REDACTED]:user/repo and then pi update <same string> to update as we patch

ReadREADME.md

Readsetup.sh

Readpackage.json

ReadREADME.md

Readpackages.md

Readextensions.md2000 lines

ReadREADME.md

Readsetup.sh

ReadSKILL.md120 lines

ReadREADME.md

Editedsetup.sh

EditedSKILL.md

EditedREADME.md

Done — I fully ported skills/creative/manim-video/ into a new pi package/extension repo and wired /manim.

✅ What I created

Repo: ~/dev/pi-manim-video
Remote: [REDACTED]:maujim/pi-manim-video.git (created via pnpm dlx gh-axi repo create ... --clone)
Pushed commit: 9542d3b

✅ What’s inside

extensions/manim-video/index.ts
- Registers /manim
- Registers resources_discover to load bundled skill:
  - skills/manim-video/SKILL.md
Full port of original skill directory:
- skills/manim-video/SKILL.md
- skills/manim-video/README.md
- skills/manim-video/references/* (all files)
- skills/manim-video/scripts/setup.sh

✅ Lightweight setup script changes

I replaced the old setup behavior with a lighter checker that:

Uses uv checks for Python + Manim executable
Checks pdflatex and ffmpeg
Does not auto-install anything
Prints manual install guidance only

✅ Installed + updated in pi

Executed successfully:

And confirmed via pi list.

If you want, I can now add a /manim-init helper command that scaffolds a fresh project folder (plan.md, script.py, concat.txt) in one shot.

out of curiosity is it possible to make it so this skill is dispatched with a specific model from a specific provider?? Let's say in the future I train a model that is fine-tuned to be really good at generating videos like this.

Yes — absolutely possible.

In pi, skills themselves are just prompt content, but your extension command (/manim) can control model selection before dispatching /skill:manim-video.

Practical pattern

In /manim handler:

Look up your target model (provider + model id)
Call await pi.setModel(targetModel)
Send pi.sendUserMessage("/skill:manim-video ...")

Optionally:

Save previous model and restore it after the run (on agent_end), or
Keep the session pinned to your video model until user changes it.

Notes

This is extension-level behavior, not SKILL.md-native behavior.
Works great once your finetuned provider/model is added (via built-in model registry or pi.registerProvider()).
You can also make this configurable (/manim-model set provider/model-id).

If you want, I can patch your pi-manim-video extension to do exactly this with a default + optional auto-restore toggle.

can we add into this manim extension that it was created by porting from the hermes skill and what commit this was done against? and also link to the specific subdirs we used? This way we can update if the skill gets updated yk

ReadREADME.md

ReadSKILL.md40 lines

EditedREADME.md

Yep — done ✅

I added explicit provenance tracking to ~/dev/pi-manim-video and pushed it.

Added

README.md now includes:
- that this was ported from Hermes
- exact upstream repo + commit:
  - e902e55b26aab4658debab070fc1048b22517158
- pinned links to:
  - skill root
  - references/
  - scripts/
UPSTREAM.md (new file) with:
- upstream snapshot metadata
- local path mapping (upstream → this package)
- update procedure for future syncs
- note that setup.sh is intentionally adapted (lightweight, no auto-install)

Pushed

Commit: 22cd475
Repo: [REDACTED]:maujim/pi-manim-video.git

If you want, I can also add a tiny /manim-upstream command that prints this provenance info in-chat so you can quickly inspect source/commit without opening files.

commit and push this. Does this skill ask the agent to run commands always insiade a fresh uv env, or to use the users env, or to at least do something like uvx manim

Committed and pushed already ✅
Latest pushed commit is:

22cd475 on master
repo: [REDACTED]:maujim/pi-manim-video.git

And right now: no uncommitted changes.

On your question: currently this skill does NOT enforce fresh uv env / uvx manim.

What it does today:

setup checker uses uv only to verify availability
skill instructions still render with plain commands like:
- manim -ql ...
- manim -qh ...
references still include some pip install ... snippets for optional voiceover tooling

So execution is effectively in the user’s existing env / PATH, unless the model decides otherwise.

If you want, I can patch it so the skill consistently prefers one strategy, e.g.:

Always uvx manim ... (tool-style ephemeral execution), or
Always uv run --no-project manim ..., or
Project-local venv workflow with explicit uv venv && uv pip install ... instructions.

prefer 1) when just running commands doing one off renders prefere 2) and 3) in more mature codebases

ReadSKILL.md80 lines

Readrendering.md220 lines

EditedSKILL.md

Editedrendering.md

Editedtroubleshooting.md

Editedproduction-quality.md

Done — committed and pushed ✅

Commit: 798b1c8
Message: docs: prefer uvx for one-offs and uv run/project env for mature workflows
Repo: [REDACTED]:maujim/pi-manim-video.git

Answer to your question (now)

Previously it didn’t enforce this clearly.
Now the skill/docs explicitly guide:

One-off renders: prefer uvx manim ...
Mature codebases: prefer uv run --no-project manim ...
Reproducible mature projects: prefer project-local uv env (venv + pinned deps)

I updated:

skills/manim-video/SKILL.md
skills/manim-video/references/rendering.md
skills/manim-video/references/troubleshooting.md
skills/manim-video/references/production-quality.md

If you want, I can also add a strict /manim-mode oneoff|mature|repro command that injects the chosen policy into the session automatically.

commit and push then update Pi

Done ✅

Repo was already committed/pushed at:
- 798b1c8 docs: prefer uvx for one-offs and uv run/project env for mature workflows
Ran:
- pi update git:[REDACTED]:maujim/pi-manim-video
Pi updated successfully to latest remote commit.

ok one more update. when the model is running the render command, teach it to pipe stdout and stderr to a file or to stream it elsewhere. I don't want all the context of the stdout video render to fill up ctx but i also wanna have errors and success to be read to keep the workflows going.

ReadSKILL.md190 lines

Readrendering.md220 lines

Readindex.ts

EditedSKILL.md

Editedrendering.md

Editedindex.ts

Done ✅ — updated, committed, pushed, and refreshed in Pi.

What changed

I taught the extension/skill to keep render logs out of LLM context by default:

skills/manim-video/SKILL.md
- Added Render Logging Policy:
  - always redirect stdout+stderr to log file
  - report only status/exit/artifact/log paths
  - include only bounded tail on failure (tail -n 120)
  - use tee only when explicit live streaming is needed
skills/manim-video/references/rendering.md
- Added LLM-safe Render Logging section with concrete command templates (uvx + log redirection, failure handling, optional tee streaming).
extensions/manim-video/index.ts
- /manim now appends a short render-policy instruction every time it dispatches /skill:manim-video, to reinforce this behavior operationally.

Git + Pi update

Commit: 0389ed9
Message: feat: enforce render log redirection to keep Manim output out of context
Pushed: origin/master
Ran:
- pi update git:[REDACTED]:maujim/pi-manim-video
Update completed successfully.