## Why
Project instructions must respect the selected environment's filesystem read
permissions. Tightening those permissions for a later turn must not allow
previously cached instructions to reach the model.
## What changed
- Apply each environment's filesystem sandbox while discovering and reading
`AGENTS.md` files.
- Fail thread or turn setup when sandboxing blocks a discovered instruction
file, while allowing a restricted project with no instructions to start.
- Clear cached instructions before refresh and include the Windows sandbox
level in the cache key.
- Ignore inaccessible ancestor marker probes so readable instructions in the
selected working directory can still load.
## Testing
Added coverage for restricted projects, denied instruction files, cache
invalidation after permissions tighten, and unreadable ancestor markers.
GitOrigin-RevId: 6ea1a27b9c873a6260e4f87d42ae1317a6a4ae4b
Adam Perry @ OpenAI
7ece061767