[debuggingfuture]

Containers run on the rpc transport by default, the R2 cache uploads each content-addressed key once and skips the save on a restore it cannot apply, and the Rust sandbox variant ships a sha256-pinned prebuilt DuckDB shared library.

rpc transport by default

• resolveSandboxTransport (packages/runtime-cf/src/sandbox-transport.ts) resolves one transport per execution: the CONFIG_KV pin sandbox.transport:<repo>, else the Worker-wide SANDBOX_TRANSPORT var, else rpc. A value that is not a transport name is skipped at its level.
• The pin overrides the var in either direction, so an operator moves one repo back to http without a deploy.
• makeCFRuntimeLive pins the resolved transport onto every container the execution acquires (sandbox-cf.ts, once per container) and hands the same value to the cache Layer. A container keeps the transport it started with for its life.
• The dispatcher passes the raw pin and var through; the precedence lives in the one pure helper, covered by sandbox-transport.test.ts.
• runs/README.md § Container transport documents the precedence, the rollback key and why rpc.

Upload each cache key once

• RestoreFn reports hit, miss or unapplied. composeRestoreOr runs onMiss on miss and unapplied, and saves only after a miss. A restore error still degrades to a miss.
• composeSave heads the key before packing. A stored archive skips tar czf and the upload, so parallel misses on one key upload once. A failed head check packs as before.
• The live restore streams an archive in on rpc. On http / websocket it reads the key's head, reports unapplied without downloading the body, and logs the transport once per execution. An rpc apply that fails cancels the unread body and reports unapplied.
• contextful-gate parts keep install: true, and a comment records why: the stage runs the consumer's contextful-ci, and the consumer's .cargo/config.toml can set net.offline, so nothing in this repo shows that the stage fetches its own crates.

Prebuilt DuckDB in the Rust sandbox variant

• The WITH_RUST variant of infra/Dockerfile.sandbox-demo (RunSandbox) ships DuckDB's prebuilt libduckdb.so and headers under /opt/duckdb, so a consumer that links DuckDB skips the C++ compile that dominates a cold cargo test.
• One ARG DUCKDB_VERSION pins the release; the libduckdb-linux-amd64.zip asset is verified with sha256sum -c against the digest GitHub publishes for it. Other architectures fail the build.
• DUCKDB_LIB_DIR / DUCKDB_INCLUDE_DIR are exported, and an /etc/ld.so.conf.d/duckdb.conf entry puts the library on the loader path for every exec user. The build asserts ldconfig -p resolves libduckdb.so.
• libduckdb-sys reads both variables only in its linked (non-bundled) build, and in a linked Linux build DUCKDB_LIB_DIR takes precedence over DUCKDB_DOWNLOAD_LIB=1. The consumer opt-in is documented on the layer and in the RunSandbox comment in wrangler.jsonc.
• The layer ships in the WITH_RUST variant rather than the lean image the ADR names: RunSandbox is the only class with a Rust toolchain to build a linked consumer, and infra/Dockerfile.sandbox sets WITH_RUST for no class.

Spec: specs/adr/0007-cost-efficient-execution.md § Upload each cache key once, § Fan-out parts install nothing their stage repeats (the contextful-gate parts keep install: true, as above) and § Prebuilt heavy native dependencies in the image; specs/adr/0008-cheaper-waiting-and-transport.md § rpc transport by default (#232). ADR-0007 § Dispatch only work is #234.

Testing: pnpm lint, pnpm typecheck and pnpm test (207 files, 2836 passed, 1 skipped) pass. sandbox-transport.test.ts covers the default, the var, the pin in both directions and skipped values. cache-r2-live.test.ts runs the live cache closures against a mocked getSandbox and a fake bucket (http, websocket and rpc restores, an rpc miss, an rpc writeFile or tar xzf failure, a save of a stored key); cache-r2.test.ts covers unapplied and composeSave. Not run: an image build of the DuckDB layer and a linked cargo build against it; no PR CI builds images.

[flaredispatch-fractalboxdev]

AI code review — 💬 Comment

Risk tier: full · 0 critical · 1 warnings · 0 suggestions

Reviewers: security ⚠️ · performance 1 · code-quality 1 · documentation 0 · release-management ⚠️ · compliance ⚠️ · agents-md ⚠️

1. ⚠️ Warning — Concurrent cache misses still duplicate packing and uploads

📍 packages/runtime-cf/src/cache-restore-or.ts:48-63

'composeSave' performs a non-atomic 'exists' check followed by 'pack'. Parallel executions can all observe a missing key, then each run the expensive tar creation and upload, contradicting the stated upload-once behavior and potentially causing substantial duplicate CPU, I/O, and R2 traffic. Add per-key coordination/locking or use an atomic create-if-absent mechanism before packing/uploading.

📋 View full logs & reviewed diff ↗

[flaredispatch-fractalboxdev]

AI code review — 🛑 Request changes

Risk tier: full · 1 critical · 2 warnings · 1 suggestions

Reviewers: security 0 · performance 2 · code-quality 0 · documentation 1 · release-management 0 · compliance 2 · agents-md 0

1. ⚠️ Warning — Every pnpm cache restore performs another install

📍 packages/core/src/primitives/install-cached.ts:154-163

The new 'relink' command runs after every restore, including cache hits and cache misses. 'pnpm install --frozen-lockfile --prefer-offline' still performs dependency resolution and filesystem work, and on a miss it repeats the install that 'onMiss' already performed. This adds avoidable latency to every pnpm run and can negate the cache benefit; use a narrowly scoped relink operation or only run it on hits when the archived tree actually requires workspace links.

2. ⚠️ Warning — Concurrent misses can still pack and upload the same cache key

📍 packages/runtime-cf/src/cache-restore-or.ts:51-63

The 'exists' check and 'pack' call are separate, non-atomic operations. Parallel executions can both observe a missing key, independently run the expensive archive creation, and upload duplicate archives, despite the comment promising upload-once behavior. Add per-key coordination or an atomic/conditional creation mechanism before packing, or accept and document the duplicate-work behavior.

3. 💡 Suggestion — Clarify that rpc is only the default transport

📍 runs/README.md:505-506

The opening sentence says containers run on the SDK's 'rpc' transport, but the documented configuration can resolve them to 'http' or 'websocket' via the repo pin or 'SANDBOX_TRANSPORT'. Reword this to say containers default to 'rpc' and use the resolved transport according to the precedence below, so readers are not given an unconditional claim that contradicts the configuration instructions.

4. 🛑 Critical — DuckDB license files are omitted from the distributed image

📍 infra/Dockerfile.sandbox-demo:171-172

The image copies only 'libduckdb.so' and three headers from the DuckDB release archive; no DuckDB license or copyright notice is retained. Since this adds and distributes DuckDB binaries/headers, preserve the applicable license notices in the image or another distributed compliance location.

📋 View full logs & reviewed diff ↗

[flaredispatch-fractalboxdev]

AI code review — 💬 Comment

Risk tier: full · 0 critical · 2 warnings · 0 suggestions

Reviewers: security ⚠️ · performance 2 · code-quality 1 · documentation ⚠️ · release-management ⚠️ · compliance ⚠️ · agents-md 0

1. ⚠️ Warning — Concurrent cache misses still duplicate packing and uploads

📍 packages/runtime-cf/src/cache-restore-or.ts:52-64

The existence check and pack are not atomic: concurrent executions can all observe 'exists=false', then each run 'pack(opts)' and upload the same content. This defeats the stated upload-once optimization and can multiply tar/stream/upload CPU, container time, and R2 bandwidth under fan-out. Add per-key coordination or an atomic conditional-write/claim mechanism before packing.

2. ⚠️ Warning — Every cached install now starts a second pnpm install

📍 packages/core/src/primitives/install-cached.ts:156-164

The relink command runs unconditionally after 'restoreOr', including cache misses where 'onMiss' has already performed the full pnpm install and populated workspace links. This adds a pnpm startup and workspace/lockfile scan to every miss, and a potentially expensive install check to every hit. Restrict relinking to successful cache restores, or use a purpose-built link-only operation that cannot re-resolve or reinstall dependencies.

📋 View full logs & reviewed diff ↗