Over the break, I started by focusing on closing the loop between Generative AI output and production-ready performance, specifically addressing the $X/hour cost risks of Cloud Run GPU instances. Then, I focused on focusing - by building a task tracking tool that started my explorations into multi-GeminiCLI orchestration.
Pixie Anim
Pixie Anim is a Rust WASM GIF optimizer. Read more about how not only is it a direct reimplementation of core optimization algorithms, but uses Gemini in a novel critique an analysis process to improve those algorithms.

Try it: https://ghchinoy.github.io/pixie-anim/
LitFlow
LitFlow, a graph renderer in Lit, has been improved with more different types of auto layouts in v0.5.
Tasker
Tasker is now more than a multi project task watcher and provides idea expansion and suggested next steps.
This is a short video of the project created after expanding an idea. The project has goals, architecture, as well as tasks in epics, prioritized.
Fieldnotes
Fieldnotes - where I'm writing this right now - has expanded into more of an editing system, with an analysis pane to provide relations between posts and will eventually be built out to suggest gaps and topics.

Docstats
A refresh of docstats was pretty insightful - a lot changes in 6 months of coding agents. Docstats will eventually be incorporated into Fieldnotes.
Use docstats to get readability scores for this PDF: https://arxiv.org/pdf/2503.05244 and explain the results to me.
Readability Analysis: arXiv 2503.05244
This document presents the readability scores for the paper 2503.05244.
Readability Summary
| Metric | Score | Interpretation |
|---|---|---|
| Flesch Reading Ease | 25.88 | Very Difficult. Typical for academic papers. |
| Flesch-Kincaid Grade | 14.57 | College Level. Requires ~15 years of formal education. |
| Gunning Fog Index | 17.59 | Graduate Level. Written for a post-graduate audience. |
| SMOG Index | 15.76 | Graduate Level. |
| Automated Readability Index | 16.82 | Graduate Level. |
| Text Standard | 15.0 | Consensus Grade Level. |
Text Statistics
- Word Count: 14,710
- Sentence Count: 743
- Syllable Count: 27,970
- Average Sentence Length: ~19.8 words
- Average Syllables per Word: ~1.9
Analysis
The scores indicate a highly technical academic document. The low Flesch Reading Ease and high Grade Level indices (Flesch-Kincaid, Gunning Fog, SMOG) are consistent with professional research papers that employ specialized terminology and complex syntactic structures.
Home Control
Home control dashboard and text user interface control created with a discovery component. Gemini CLI + letting it go to town on Linux network tooling.

Riptide
Riptide is more stable now and Gemini CLI is very helpful in discovering where Gemini 2.5 Computer Use is either too strict (i.e. not starting a request with "navigate to.." or similar will try to use the (non existent) search too), or too vague.
Here's riptide turning off the lights.

Banana Mesh
Create 3D Meshes using Nano Banana Pro and the Hunyuan3D-2 on Cloud Run GPU - an updated version of the excellent example by Nikolai Danylchyk.
Runner ups
Banana Peel
Integrating Qwen Layers with Nano Banana Pro. Lots of issues with Cloud Run GPU limitations for this large model. As you can see, it "works."

Shepherd 🐑
A tool to monitor Cloud Run GPU instances and auto shut them down. This was borne out of the need to not have GPUs run overnight (see Banana Peel and 3D Meshes, above), so I have a per-service configurable scheduled check and quiescence job (Cloud Run jobs) with a dashboard.

Upcoming
- Idea Factory - this is a stand alone component pulled out of Tasker to just capture and auto-elaborate on ideas. For this one, I have a prototype with the new Gemini Interactions API with Deep Research and am exploring Stitch again for the UI design.
- Multitasker - Tasker grouped projects that communicate with beads to start off the "year of the agent harnesses"
2026 Agent orchestrator wars begin
— ghchinoy (@ghchinoy) January 5, 2026
For those of you that remember and have scars from the container orchestration wars, welcome to the start of agent harness cage matches