Signal Core

Signal Core compresses context before agents read your code. It sends structural skeletons (signatures, types, imports) instead of full file dumps. Savings depend on how your files are shaped: small stubs save less; large source files with long bodies save more.

At a glance

• tree-sitter parsing extracts structure on the read path (8+ languages)
• signal_core = true in pytxo.toml turns scaffolding on for task paths
• signal_fidelity (low | medium | high) controls how much structure per task
• Savings are recorded per run and readable through the CLI and the run store
• Pytxo Desktop does not display a per-run savings figure, because the orchestrator records scaffold bytes rather than billable tokens

Measured evidence

On 2026-07-31, real-repo-signal.ps1 measured 82.9% weighted scaffold-byte reduction across 185 tracked production Rust, TypeScript, and JavaScript files in the Pytxo monorepo (1,253,675 bytes to 213,847 bytes).

This measures the bytes emitted by the structural scaffold, not billable model tokens, task success, cost, or end-to-end agent quality. Re-run the benchmark on your own repository; file shape and language mix materially change the result.

Configuration

signal_core = true
signal_fidelity = "low"   # default; per-task override in [[task]]

Signal fidelity is a context tier, separate from permission tiers.

MCP reads

Tools pytxo_read and pytxo_read_scaffolded return skeletons instead of raw files when Signal Core is on. Useful from Cursor without dumping a full checkout.

Back: Three moats · Pytxo Desktop