When charmbracelet/crush merged Amolith's commit 6242e4f4, it raised the operating question now moving through agent runtimes: which side of the context border did this instruction, path, or project copy come from?
Facts
- Crush now reads
CRUSH.mdand genericAGENTS.mdfiles from user config paths and exposesglobal_context_pathsin configuration. - OpenAI Codex migrated sandbox current-working-directory inputs to
PathUri, keeping URI semantics until the receiving host validates them. - OpenClaw preserved memory prompt registration in its Codex runtime plugin, while OpenCode refactored project-copy handling around project directory strategy.
Evidence
The receipts are Crush commit 6242e4f4, Codex commit 52a50aec, OpenClaw commit 7387083a, and OpenCode commit c2e6b180.
Context
Reporter note: this article groups independent patches around one analytics view of agent work. User preference files, project-specific context, sandbox working directories, registered memory prompts, and copied project trees carry different risk and lifetime assumptions.
This continues the trust-boundary arc from Friday's credential story, but the boundary has moved inward: builders now have to ask which instructions, paths, and project copies become part of an agent run.
Limits
These are independent patches, not one shared design. The evidence shows local context being routed more carefully; it does not prove that every agent has solved stale memory, unsafe path handling, or user/project instruction conflicts.