History

2026-04-30: OpenCode

• Vibe-coding (testing skills) with limited (free) models
• Permission system
• Data: skills.sh

2026-04-22: Experiment with OpenClaw

• Play with OpenClaw
• Discover model providers Zen-Go and OpenRouter.
• Data: Pinchbench model benchmark for OpenClaw.

2026-03-26:

• Agent Frameworks for Software development.
• OpenAgentsControl seems to be my best fit.

2026-03-25: Learn Context window management.

• Takeaways/LL:
  • Small Language Models (LLMs). Model compactation techniques.
  • U-shaped reliability
• Data:
  • models.dev for model limit info.
  • Thresholds: 100Ktokens, 50%

2026-03-24: Experiment:

• Data:
  • Briefly explainig repossync burned 25.8 KTokens (13% used).
  • Explaining test_repossync took the count to 37.8 Ktokens (19% used).
  • Skillmarket

2026-03-23: Experiment:

• Install dockerized opencode. Connect a backend.
• Data: Init with model:BigPickle burned 61K tokens (30% used) and generated an AGENTS.md of 6.8kB, 877 words and 260 lines.

2026-03- : Update AI assistants comparison

2026-03-19:

• Might the limit of context window be the case for unexpected misbehavior? Try from fresh chats? => Standardize chat management.

2026-01-24: Experiment with Coding Assistants:

• Try OpenCode
• Try GroqCloud as backend AI service for FLOSS dev, as recommended by Gemini. Token off:2026Mar25.

2025-11-28: Experiment with Gratis AI online services:

• Try Claude and find out the free version is lmited to only 1 project.
• Try DeepSeek and find out that it is "disconnected" from the internet and reads a max of 10 pages per query "Disconnected" but "reading"?
• As expected, DeepSeek misconcieves some aspects.

2025-11-29: Experiment with Gratis AI online services:

• Try NotebookML becasue it is supposed to ingest a lot and dive deep.

2025-11-20: Dream of

• Presenting Vibe Debian Packaging (VDP) in DebConf. Gen Z might preffer to learn by guided doing. VDP might open the door for non-tech users.
• VDP needs reproducibility => Evolve the process of training an AI as assisstant on a git repository for reproducibility and contributions.
• Personal branch naming convention: LucIA, ValerIA, MarIA as Intelligent Assisstant (IA) names (repo branches).

Before....: Research and discover effective usage:

• Use Gemini at work. Tell the kind of tasks that work apat from the ones that seem to but don't.
• Try Chat GPT (de-Google), DeepSeek (for local exec), and Duck AI (for privacy).
• Try to run DeepSeek locally with (dockerized) Ollama. The laptop's cooler goes wild.