1.4.0

[cybaea]

v1.4.0: The "Analyst" Update 🧠 + 🧮

This isn't just a chat update—it's a brain transplant.

In version 1.4.0, we are introducing the Computational Solver, a specialized sub-agent that gives your vault the ability to write and execute Python code.

Your Agent is no longer limited to summarizing text. It can now act as a Data Analyst:

• Analyze Data: "Read my @Expenses note and forecast next month's spend."
• Crunch Numbers: "Calculate the average weekly change in my @WeightLog."
• Solve Logic: Handle complex math and logic puzzles that stump standard LLMs.

We have also rebuilt the engine under the hood. The new "Greedy Packing" strategy replaces arbitrary text limits with intelligent context management, allowing the Agent to read entire documents (up to your token budget) to capture the full nuance of your long-form writing.

From passive storage to active reasoning—enjoy your new analyst.

(See our newly published Roadmap for where we are going next!)

🚀 Added

• Computational Solver: The agent can now write and execute Python code to analyze data from your vault (e.g., "Read @Monthly Expenses and forecast next month's spend").
• Settings: New Enable code execution toggle for the Computational Solver above(Default: Off) and corresponding Code model selector in the Models tab.
• Settings: New Context window budget setting (Default: 200,000 tokens). This allows users to control how much "memory" the agent uses per request, balancing deep context against API rate limits.

⚡ Changed

• Context Engine: Replaced the static 50,000-character limit per note with a "Greedy Packing" strategy. The agent now reads entire documents whenever the token budget allows, significantly improving its ability to understand long reports and avoiding arbitrary cut-offs.

🛠 Developers

• Dynamic Tools: Updated AgentService to dynamically construct the tools array at runtime. This resolves an issue where conditional tools (like the solver) were defined but not correctly passed to the model if disabled.
• Response Parsing: Refactored GeminiService to manually parse multi-part responses (text, executableCode, codeExecutionResult), resolving SDK warnings about mixed content types.
• Context Limits: Removed hardcoded MAX_TOTAL_CONTEXT constants. Context limits are now calculated dynamically based on the user's contextWindowTokens setting.
• System Prompt: Updated the default system prompt to include specific instructions on when and how to utilize the computational_solver tool.

---

This discussion was created from the release 1.4.0.