Function build_cache_note 

pub fn build_cache_note(node: &str, allow_write: bool) -> String

What a seat is told about the shared build cache — one of two notes, chosen by whether the seat may write at all.

Spliced into every node prompt (in [crate::graph::wave] and crate::graph::Runner::synthesize_brief) when the run’s config declares a CARGO_TARGET_DIR — which is also the directory the verify commands build into. The text is stable so tests can assert on it; the value of the variable is not spelled out because a write-allowed seat reads it from its own environment, and a prompt that hardcodes a path would go stale the moment the config moves the cache.

allow_write must agree with whether the caller actually hands the seat CARGO_TARGET_DIR (see crate::agent::Invocation::cache_dir) — a read-only seat that is still told “build through it” is exactly how a sandboxed reviewer’s write refusal to a directory it was never meant to touch got reported as a defect in the patch under review. So a read-only seat is told plainly that it has no shared cache and that a write refusal anywhere outside its own worktree is expected, not evidence of anything.

The fund-transfer reality the write-allowed note exists to prevent: an implementer that builds with its own CARGO_TARGET_DIR (or lets cargo create a fresh target/ in the worktree) is compiling a second copy of the world that nobody prunes, on a machine that has already had that exact failure once. It also spells out the one thing a test name filter cannot do — cargo test report:: still compiles every integration target in the workspace, because the filter selects which tests run, not which targets get built — so a seat asked for a narrow check knows to reach for --lib/--test instead of assuming a filter alone bounds the build.

node is the graph node this is spliced into ("review", "fix", …). A reviewer or fixer gets an extra paragraph saying full verification is magi’s own job, not theirs to repeat — the same duplicated-full-suite cost neither note’s own advice does anything to prevent on its own, since a seat that dutifully stays inside its own worktree can still spend the round re-running the whole suite there. Phrased as a request, not a guarantee: magi has no way to stop a seat from running cargo test --all-targets anyway, so the note asks rather than claims it enforces anything.