CoreAI v6.7.0

Prompt-delivery fix, a measured Lua guard optimization, and a full read-only audit of the Hub, core and Unity packages.

Fixed

• Built-in agent prompts never reached a Unity host. AgentPromptsInstaller chains the Resources/AgentPrompts/System provider ahead of the built-in one, so a prompt shipped there permanently shadows its C# const — and the package shipped copies of seven built-in prompts. The shipped Programmer.txt had already drifted: it was missing the "answer plain questions directly — do NOT call read_skill or any tool for those" rule, the report()/logic_* globals list, and the Forbidden: io, os, require, load, loadfile, dofile, debug line. Every Unity host was therefore running an older Programmer prompt than the code said.

  The eight shadowing/dead assets are deleted (PlayerChat.txt was dead outright — the role id became PlainChat), leaving DeveloperSampleAgent.txt, which has no const. Resources/AgentPrompts/System is now purely the consumer's override slot, as its own doc comment says. A new test, NoBuiltInRolePromptIsShadowedByAShippedResourceCopy, fails if such a copy is reintroduced.

Changed

• Lua execution guard: the wall clock is now sampled every 64th hook fire instead of every fire. Measured: Stopwatch.GetTimestamp() ~44 ns vs ~14 ns for the heap read, i.e. ~76% of the work inside a hook that fires every 4 VM instructions. The step budget is still charged on every fire, so a runaway is cut exactly as before. Guarded-execution overhead: 3.67× → 3.21×.

Docs

• dev-docs/LUA_PERF_AUDIT_v6.6_2026-07.md — two findings worth reading: the published Luau comparison was measured without the execution guard and understates the real gap by ~3.7×; and the guard's cost is dominated by the number of hook fires (~600 ns each — the VM's async hook dispatch), not the work inside the hook. Raising the batch to 64 measures 1.21× overhead (a 2.65× speedup) but was deliberately not shipped: the small batch is load-bearing for the allocation-bomb backstop. The doc specifies the adaptive-batch design that would capture the win safely.
• dev-docs/CODE_AUDIT_v6.6_2026-07.md — audit backlog for CoreAIHub, CoreAI and CoreAiUnity.

Full EditMode suite: 2386 passed, 0 failed, 9 skipped (intentional).