bench/package.json pins @tangle-network/agent-eval at ^0.114.0 (lockfile: 0.114.0) while the repo root declares ^0.115.1 (peer range >=0.115.1 <1.0.0).
Today bench mostly gets away with it because its agent-eval imports are type-only, but any bench script that runtime-imports agent-eval executes against a stale contract, one version behind what the runtime itself is built for. The drift was first noticed during the live self-improvement campaign (script: agent-runtime-swe commit 3b93916) at 0.106.3-vs-0.108.1 and has recurred at 0.114.0-vs-0.115.1 — it regenerates on every root bump.
Align the pin now and prefer a single source for the version (workspace protocol or a sync check in CI) so bench/ can't drift behind root again.
bench/package.jsonpins@tangle-network/agent-evalat^0.114.0(lockfile: 0.114.0) while the repo root declares^0.115.1(peer range>=0.115.1 <1.0.0).Today bench mostly gets away with it because its agent-eval imports are type-only, but any bench script that runtime-imports agent-eval executes against a stale contract, one version behind what the runtime itself is built for. The drift was first noticed during the live self-improvement campaign (script: agent-runtime-swe commit 3b93916) at 0.106.3-vs-0.108.1 and has recurred at 0.114.0-vs-0.115.1 — it regenerates on every root bump.
Align the pin now and prefer a single source for the version (workspace protocol or a sync check in CI) so bench/ can't drift behind root again.