Context selection vs runtime control: can we design one falsifiable OpenCode experiment?

[feiai2026]

OpenAgentsControl is the closest public implementation we have found to the product direction we are testing: ContextScout, ContextManager, the MVI principle, and integrations for OpenCode and Claude Code.

Our narrower hypothesis is that a coding agent needs a model-driven compiler before execution: select project instructions, decisions, state, skills, tools, permissions, history, and file evidence for each task, then prove exactly what entered every provider request.

We have an encouraging but deliberately limited signal. Across 10 task types, 30 paired trials, and 66 real model calls, manually frozen minimal Context Packages reduced total tokens in 30/30 pairs, averaging -50.8%, while irrelevant/forbidden-information leakage fell from 23.3% to 6.7%. This does not prove that an automatic compiler improves task success.

Rather than build a duplicate of OAC, I would like to test whether the two layers are complementary:

1. Can the current OAC path prove that unselected native instructions, skills, tools, MCP definitions, permissions, history, and tool results never bypass context selection?
2. What raw experiment supports the public token-reduction result, and which task-quality metric should accompany it?
3. Could we run one shared OpenCode benchmark with isolated workspaces and per-request traces, comparing native OpenCode, OAC, AIOS-style selection, and a combined path?

Brief and current evidence limits: https://feiai2026.github.io/aios-context-compiler/

A critical answer such as “this belongs inside OAC” is welcome. If the boundary is genuinely complementary, a short design review or a small integration spike would be a useful first collaboration.