July 2026 · Product method
AI coding context health: hooks, MCP tools, and fresh sessions
Your coding agent can carry more context than the visible chat suggests. Skill descriptions may be discoverable, an MCP server may be configured or explicitly requested as always-loaded, lifecycle hooks can inject instructions, and a long session can keep accumulating history. Those are different states, so one generic “context used” number is not enough.
The missing distinction: available is not active
A skill listed to the model is not the same as a skill explicitly invoked. An MCP server in configuration is not proof that one of its tools ran. A lifecycle hook is different again: events such as SessionStart, UserPromptSubmit, or SubagentStart can add runtime context without appearing as a normal tool call.
aibill separates discoverable, explicitly invoked, MCP-configured, explicit always-load requests, hook-injected, and invocation-unobservable states. MCP configuration proves availability or intent—not the schema payload loaded at runtime. It reads installed hook configuration as metadata, but never executes the command and never guesses the emitted token payload.
One decision, backed by your own history
Context Health prioritizes directly observed compaction evidence. Where the transcript exposes it, it otherwise compares latest-turn input context with comparable prior sessions from the same coding agent and project—not cumulative lifetime totals. The action may be to preserve a checkpoint and start fresh, inspect configured hooks or inventory with no matching invocation, continue, or collect more history.
That is a workflow signal, not a universal efficiency claim. Token volume does not prove code quality, latency, or money saved, and a configured item with no matching invocation may still be valuable tomorrow.
Run it in the interface you already use
The terminal JSON, MCP tool, and Glance card consume the same versioned contract. That matters more than visual consistency: a session should not be “healthy” in one interface and “start fresh” in another.
What we borrowed—and what we did not
Developer tools such as Ponytail show the appeal of a memorable single job, operational skills, portable adapters, and a public benchmark method. aibill adopted those product-engineering ideas. It did not copy always-on prompt injection, reuse another project's benchmark percentages, or treat lines of code as a universal cost proxy.
Our public fixture benchmark tests classification, decision precedence, and safety invariants. It explicitly does not claim universal token savings, faster delivery, or better code. Those claims would require a controlled baseline and raw results.
Privacy depends on the surface
Terminal and Glance stay on the machine and aibill sends no telemetry. If you explicitly invoke an MCP-backed skill, the selected structured result is returned to that AI client and is governed by the client's data policy. That boundary is more useful than a vague promise: you can choose the terminal when you want no AI-client handoff and the plugin when conversational explanation is worth it.